Introduction#
ZDTaichu5.0-9B is a multimodal foundation model for general visual understanding, spatial reasoning, agentic tool use, and embodied-AI research. It combines a Qwen3.5-9B language backbone with a C-RADIOv4-H vision encoder, supports text, images and videos with any-resolution visual input.
Within the 9B-scale general-purpose VLMs compared in this release blog, ZDTaichu5.0-9B retains first-tier general visual understanding while supporting spatial reasoning, high-level embodied VLM reasoning, and agent tasks under the reported evaluation settings. Rather than trading broad visual competence for specialization, it layers a more comprehensive spatial, embodied, and agent capability profile on top of a strong general-vision foundation.
The model accepts text, one or more images, and video. It is designed for:
- general image, document, chart, diagram, and OCR understanding;
- visual mathematics and knowledge-grounded visual question answering;
- fine-grained 2D relations, multi-view association, 3D scene understanding, perspective taking, and mental transformation;
- multi-step and multi-turn tool use;
- spatial perception, affordance understanding, and planning for VLA and embodied-AI adaptation.
Highlights#
- Strong general vision and broad capabilities: remains in the leading group of 9B-scale general-purpose VLMs across images, documents, charts, diagrams, OCR, visual mathematics, multiple images and video, while extending to spatial reasoning, high-level embodied understanding and multi-step agent tasks.
- Leading spatial reasoning and embodied understanding: leads spatial capability among the compared 9B-scale general-purpose VLMs, with strong results on SparBench, ViewSpatial, MMSI-Bench and MindCube-tiny. Scores of 48 on ERQA and 56 on RoboSpatial cover scene reasoning, affordances and interaction-oriented understanding.
- Strongest agent capability among the compared 9B-scale general-purpose VLMs: leads the reported TAU2-Bench (87.7) and Claw-Eval (71.4) comparisons, and reaches 93.7 on IFEval.
- Entropy-Gated Adaptive Recurrent Reasoning: Dynamically allocates additional recurrent refinement steps in latent space to more challenging tokens, enabling greater computational depth where needed and improving reasoning performance on complex tasks.
Model Overview#
| Item | Specification |
|---|---|
| Model type | Multimodal causal language model with vision encoder |
| Language backbone | Qwen3.5-9B LLM Decoder |
| Vision backbone | C-RADIOv4-H |
| Context length | Up to 128K tokens |
| Vision resolution | Any-resolution visual input |
| Input modalities | Text, single image, multiple images, and video |
Adaptive recurrent reasoning#
ZDTaichu5.0-9B uses entropy gating to allocate inference computation dynamically. The uncertainty of the current token determines whether to further refine hidden states in latent space, adapting reasoning depth during generation.
- Trigger recurrence by uncertainty: After a standard forward pass, output entropy measures uncertainty for the current token. Low-entropy tokens proceed directly to output, while high-entropy tokens trigger repeated computation through an intermediate layer block.
- Refine states in latent space: Damped, anchored updates progressively refine hidden states, allocating additional computation to difficult tokens.
- Stop using convergence signals: KL divergence and hidden-state residuals guide stopping, dynamically controlling the depth of recurrent computation.
- Select an output from the trajectory: Trajectory readout selects the lowest-risk state, with rollback when needed.
Benchmark Results#
Comparison with open models
Comparison with closed models
The two figures compare ZDTaichu5.0-9B with open and closed models across general visual understanding, spatial and embodied capabilities, and agent and text capabilities.
Spatial and embodied reasoning
| Area | Benchmark | ZDTaichu5.0-9B | Qwen3.5-9B | STEP3-VL-10B | gemma4-8B-E4B | Gemini 3 Pro | Grok 4 | GPT-5.2 |
|---|---|---|---|---|---|---|---|---|
| Basic spatial perception | CV-Bench | 86.82 | 87.19 | 83.49 | 68.10 | 90.07 | — | 86.84 |
| 3DSRBench | 60.96 | 56.78 | 55.01 | 53.62 | 68.92 | 54.93 | 60.20 | |
| SparBench | 51.82 | 50.79 | 45.68 | 28.50 | 48.74 | 44.76 | 55.07 | |
| Complex spatial reasoning | ViewSpatial | 62.50 | 48.20 | 46.14 | 41.68 | 50.36 | 43.23 | 47.30 |
| MMSI-Bench | 47.20 | 38.70 | 32.18 | 29.20 | 45.20 | 37.80 | 41.30 | |
| MindCube-tiny | 78.27 | 57.60 | 62.81 | 48.85 | 70.87 | 63.56 | 60.38 | |
| Embodied interaction | ERQA | 48.00 | 41.50 | 47.75 | 30.20 | 66.00 | — | 59.80 |
| RoboSpatial | 56.00 | 54.10 | 52.86 | 49.43 | 57.40 | — | 43.78 | |
| VSI-Bench | 59.69 | 55.68 | 42.42 | 32.91 | 52.51 | 47.92 | 54.49 |
Scroll to compare all models ↔
General visual understanding
| Area | Benchmark | ZDTaichu5.0-9B | Qwen3.5-9B | STEP3-VL-10B | gemma4-8B-E4B | Gemini 3 Pro | Grok 4 | GPT-5.2 |
|---|---|---|---|---|---|---|---|---|
| Multi modal Reasoning | MathVista Mini | 84.50 | 85.70 | 83.97 | 65.30 | 87.90 | 72.50 | 83.10 |
| WeMath | 75.90 | 75.20 | 73.03 | 50.19 | 86.90 | — | 79.00 | |
| MathVerse Mini Vision Only | 76.40 | 84.14 | 74.60 | 53.55 | — | — | — | |
| General VQA | MMStar | 76.80 | 79.70 | 77.48 | 62.00 | 83.10 | 69.60 | 77.10 |
| AI2D | 91.48 | 90.20 | 89.35 | 79.15 | 94.10 | — | 92.20 | |
| RealWorldQA | 76.99 | 80.30 | 74.44 | 59.08 | 83.30 | — | 83.30 | |
| OCR | OCRBench | 85.50 | 89.20 | 86.75 | 76.90 | 90.40 | — | 80.70 |
Scroll to compare all models ↔
Language, reasoning, and agents
| Area | Benchmark | ZDTaichu5.0-9B | Qwen3.5-9B | STEP3-VL-10B | gemma4-8B-E4B | Gemini 3 Pro | Grok 4 | GPT-5.2 |
|---|---|---|---|---|---|---|---|---|
| Knowledge | MMLU-Pro | 77.20 | 82.50 | 76.02 | 69.40 | 89.80 | 85.90 | 87.40 |
| MMLU-Redux | 88.40 | 91.10 | 86.50 | 85.30 | 95.90 | 86.22 | 95.00 | |
| Instruction following | IFEval | 93.70 | 88.72 | 82.16 | 87.80 | 93.50 | 92.80 | 94.80 |
| IFBench | 69.00 | 64.50 | 41.49 | 34.70 | 70.40 | 53.70 | 75.40 | |
| Reasoning and coding | AIME 2025 | 86.70 | 83.75 | 87.66 | 41.30 | 95.00 | 91.70 | 100.00 |
| AIME 2026 | 89.20 | 87.92 | 88.75 | 42.50 | 90.60 | — | 96.70 | |
| HMMT Feb 2025 | 84.20 | 83.20 | 78.18 | 26.70 | 97.30 | 90.00 | 99.40 | |
| HMMT Feb 2026 | 72.70 | 73.48 | 63.64 | 33.70 | 86.36 | — | 96.97 | |
| LiveCodeBench v6 | 73.40 | 65.60 | 58.86 | 52.00 | 90.70 | — | 87.70 | |
| General agent | TAU2-Bench† | 87.70 | 79.10 | 81.70 | 42.40 | 85.40 | — | 87.10 |
| Claw-Evalgeneral Avg† | 71.40 | 66.50 | 66.60 | 52.10 | — | — | — |
Scroll to compare all models ↔
Bold indicates the best score among the listed open-source models; underlining indicates the best score among all listed models. Scores leading both comparisons are both bold and underlined. Tied best scores receive the same marking. Missing scores are excluded from the comparison.
† Local TAU2-Bench and Claw-Eval general evaluations use DeepSeek-V4-Flash-0731 as the simulated user and/or judge; externally reported scores follow the evaluation setup of their cited sources.
‡ Publicly reported external score. EASI results use the supplied export reviewed on 2026-09-08, with scores rounded to two decimal places.
For multi-image spatial reasoning evaluations such as ViewSpatial, MMSI-Bench, MindCube-tiny, and VSI-Bench, the following output-format requirement was added to the evaluation prompt: You FIRST think about the reasoning process as an internal monologue and then provide the final answer. The reasoning process MUST BE enclosed within <think> </think> tags. The final answer MUST BE put in \boxed{}.
Capabilities and demonstrations#
Explore general vision, spatial reasoning, multi-image and video understanding, agentic tool use, and embodied planning through concrete tasks.
General visual understanding
Recognize objects, attributes and scenes, read text in images and documents, interpret charts and diagrams, and combine visual evidence with calculation and question answering.
Chart understanding and visual reasoning
Read the relationships in a function plot to identify the location of a local maximum.
ZDTaichu5.0-9B
The graph shows a 3D surface plot of the function $f(u_1, u_2)$. To find the local maximum, we need to identify the highest point on the surface.
- Identify the axes: The horizontal axes represent the variables $u_1$ and $u_2$, and the vertical axis represents the value of the function $f(u_1, u_2)$.
- Locate the peak: The surface is shaped like a pyramid. The highest point of the pyramid corresponds to the maximum value of the function.
- Determine the coordinates: Looking at the base of the pyramid, the peak is located directly above the corner where both $u_1$ and $u_2$ are at their minimum values. The axis for $u_1$ starts at 0 and goes to 1, and the axis for $u_2$ starts at 0 and goes to 1. The highest point is above the origin $(0, 0)$. As we move away from this point towards $u_1=1$ or $u_2=1$, the height of the surface decreases.
Therefore, the function reaches its maximum value at the point $(u_1, u_2) = (0, 0)$.
A
View original response
The graph shows a 3D surface plot of the function $f(u_1, u_2)$. To find the local maximum, we need to identify the highest point on the surface.
1. **Identify the axes:** The horizontal axes represent the variables $u_1$ and $u_2$, and the vertical axis represents the value of the function $f(u_1, u_2)$.
2. **Locate the peak:** The surface is shaped like a pyramid. The highest point of the pyramid corresponds to the maximum value of the function.
3. **Determine the coordinates:** Looking at the base of the pyramid, the peak is located directly above the corner where both $u_1$ and $u_2$ are at their minimum values. The axis for $u_1$ starts at 0 and goes to 1, and the axis for $u_2$ starts at 0 and goes to 1. The highest point is above the origin $(0, 0)$. As we move away from this point towards $u_1=1$ or $u_2=1$, the height of the surface decreases.
Therefore, the function reaches its maximum value at the point $(u_1, u_2) = (0, 0)$.
A
Qwen3.5-9B
The graph shows a 3D surface plot of a function $f(u_1, u_2)$.
- The vertical axis represents the value of the function $f(u_1, u_2)$. It ranges from 0 to 1.
- The horizontal plane represents the domain of the function with axes $u_1$ and $u_2$.
- Looking at the base of the plot (the $u_1, u_2$ plane), we can identify the four corners of the unit square domain $[0,1] \times [0,1]$.
- The origin corner is where the vertical axis $f$ starts. Here, $f(0,0) = 0$.
- The corner to the right (along the $u_1$ axis) has $u_1=1, u_2=0$. The surface height is low, near 0.
- The corner to the back (along the $u_2$ axis) has $u_1=0, u_2=1$. The surface height rises significantly.
- The far back corner has $u_1=1, u_2=1$. The surface height is low, near 0.
- The highest point on the surface (the peak) is located at the back-left corner of the base domain.
- Tracing the coordinates of this peak:
- It lies on the plane $u_1 = 0$ (the left edge).
- It lies on the plane $u_2 = 1$ (the back edge).
- Therefore, the function reaches its local (and global) maximum at $(u_1, u_2) = (0, 1)$.
- The value of the function at this point is 1.
This corresponds to option (C).
C
View original response
The graph shows a 3D surface plot of a function $f(u_1, u_2)$.
- The vertical axis represents the value of the function $f(u_1, u_2)$. It ranges from 0 to 1.
- The horizontal plane represents the domain of the function with axes $u_1$ and $u_2$.
- Looking at the base of the plot (the $u_1, u_2$ plane), we can identify the four corners of the unit square domain $[0,1] \times [0,1]$.
- The origin corner is where the vertical axis $f$ starts. Here, $f(0,0) = 0$.
- The corner to the right (along the $u_1$ axis) has $u_1=1, u_2=0$. The surface height is low, near 0.
- The corner to the back (along the $u_2$ axis) has $u_1=0, u_2=1$. The surface height rises significantly.
- The far back corner has $u_1=1, u_2=1$. The surface height is low, near 0.
- The highest point on the surface (the peak) is located at the back-left corner of the base domain.
- Tracing the coordinates of this peak:
- It lies on the plane $u_1 = 0$ (the left edge).
- It lies on the plane $u_2 = 1$ (the back edge).
- Therefore, the function reaches its local (and global) maximum at $(u_1, u_2) = (0, 1)$.
- The value of the function at this point is 1.
This corresponds to option (C).
C
STEP3-VL-10B
To determine where the function ( f(u_1, u_2) ) reaches its local maximum, analyze the 3D plot:
- Option (A) ((u_1, u_2) = (0, 0)): This is the front - left corner of the domain. On the plot, ( f(0, 0) = 0 ), so it is not the maximum.
- Option (B) ((u_1, u_2) = (1, 0)): This is the front - right corner. ( f(1, 0) = 0 ), not the maximum.
- Option (C) ((u_1, u_2) = (0, 1)): This is the back - left corner. The surface reaches its highest vertical value (maximum ( f )) at this point.
- Option (D) ((u_1, u_2) = (1, 1)): This is the back - right corner. ( f(1, 1) = 0 ), which is a minimum, not a maximum.
Thus, the function reaches its local maximum at ((u_1, u_2) = (0, 1)), so the answer is C.
View original response
To determine where the function \( f(u_1, u_2) \) reaches its local maximum, analyze the 3D plot:
- **Option (A) \((u_1, u_2) = (0, 0)\):** This is the front - left corner of the domain. On the plot, \( f(0, 0) = 0 \), so it is not the maximum.
- **Option (B) \((u_1, u_2) = (1, 0)\):** This is the front - right corner. \( f(1, 0) = 0 \), not the maximum.
- **Option (C) \((u_1, u_2) = (0, 1)\):** This is the back - left corner. The surface reaches its highest vertical value (maximum \( f \)) at this point.
- **Option (D) \((u_1, u_2) = (1, 1)\):** This is the back - right corner. \( f(1, 1) = 0 \), which is a minimum, not a maximum.
Thus, the function reaches its local maximum at \((u_1, u_2) = (0, 1)\), so the answer is C.
Ground truth
A
View the original prompt
Question
Hint: Please answer the question and provide the correct option letter, e.g., A, B, C, D, at the end.
Question: When does the function reach its local maximum?
Choices:
(A) (u1, u2) = (0, 0)
(B) (u1, u2) = (1, 0)
(C) (u1, u2) = (0, 1)
(D) (u1, u2) = (1, 1)
Ground truth
A
| Model | Answer | Result |
|---|---|---|
| ZDTaichu5.0-9B | A. (u1, u2) = (0, 0) | ✓ Correct |
| Qwen3.5-9B | C. (u1, u2) = (0, 1) | ✗ Incorrect |
| STEP3-VL-10B | C. (u1, u2) = (0, 1) | ✗ Incorrect |
Visual mathematics
Extract ages from the image and calculate the difference between them.
ZDTaichu5.0-9B
The two people in the image are Winston Churchill (left) and Charles de Gaulle (right).
- Winston Churchill was born on November 30, 1874.
- Charles de Gaulle was born on November 22, 1890.
The age gap between them is 1890 - 1874 = 16 years.
16
View original response
The two people in the image are Winston Churchill (left) and Charles de Gaulle (right).
* Winston Churchill was born on November 30, 1874.
* Charles de Gaulle was born on November 22, 1890.
The age gap between them is 1890 - 1874 = 16 years.
16
Qwen3.5-9B
The two people in the image are Winston Churchill (born November 30, 1874) and Bernard Law Montgomery (born November 17, 1887).
The age gap is calculated as: 1887 - 1874 = 13 years.
13
View original response
The two people in the image are **Winston Churchill** (born November 30, 1874) and **Bernard Law Montgomery** (born November 17, 1887).
The age gap is calculated as: 1887 - 1874 = 13 years.
13
STEP3-VL-10B
To determine the age gap, first identify the individuals: the person on the left is Winston Churchill (born 1874) and the person on the right is Charles de Gaulle (born 1886). Calculate the difference: ( 1886 - 1874 = 12 ).
12
View original response
To determine the age gap, first identify the individuals: the person on the left is Winston Churchill (born 1874) and the person on the right is Charles de Gaulle (born 1886). Calculate the difference: \( 1886 - 1874 = 12 \).
12
Ground truth
16
View the original prompt
Question
Hint: Please answer the question requiring an integer answer and provide the final value, e.g., 1, 2, 3, at the end.
Question: What is the age gap between these two people in image? (Unit: years)
Ground truth
16
| Model | Answer | Result |
|---|---|---|
| ZDTaichu5.0-9B | 16. years | ✓ Correct |
| Qwen3.5-9B | 13. years | ✗ Incorrect |
| STEP3-VL-10B | 12. years | ✗ Incorrect |
Reasoning in a real-world scene
Use the road layout and traffic-light position to identify the intersection it controls.
ZDTaichu5.0-9B
The traffic light is located further down the road, not at the immediate intersection in front of the camera. Therefore, it is not controlling the intersection closest to us.
B. No
View original response
The traffic light is located further down the road, not at the immediate intersection in front of the camera. Therefore, it is not controlling the intersection closest to us.
B. No
Qwen3.5-9B
A. Yes
View original response
A. Yes
STEP3-VL-10B
Based on the image, the traffic light visible in the distance is positioned at the location of the first major intersection (indicated by the pedestrian crossing and stop lines on the road ahead). There are no other intersections or traffic lights closer to the viewer's position on the road. Therefore, the traffic light ahead is indeed controlling the intersection closest to us.
A. Yes
View original response
Based on the image, the traffic light visible in the distance is positioned at the location of the first major intersection (indicated by the pedestrian crossing and stop lines on the road ahead). There are no other intersections or traffic lights closer to the viewer's position on the road. Therefore, the traffic light ahead is indeed controlling the intersection closest to us.
A. Yes
Ground truth
B
View the original prompt
Question
Is the traffic light ahead controlling the intersection closest to us?
Options
A. Yes
B. No
Ground truth
B
| Model | Answer | Result |
|---|---|---|
| ZDTaichu5.0-9B | B. No | ✓ Correct |
| Qwen3.5-9B | A. Yes | ✗ Incorrect |
| STEP3-VL-10B | A. Yes | ✗ Incorrect |
Fine-grained recognition
Locate the target animal within a complex scene and identify its color.
ZDTaichu5.0-9B
The dog is located on the beach near the group of people sitting down on the left side of the image. It has a coat that is clearly black and white.
D
View original response
The dog is located on the beach near the group of people sitting down on the left side of the image. It has a coat that is clearly black and white.
D
Qwen3.5-9B
The dog is visible on the beach near the people sitting on the sand. It appears to be a small, light-colored dog, likely gray or white. Since "white" is not an option, and the dog is not distinctly brown or yellow, and doesn't show clear black patches to suggest "black and white" (though it's small and blurry), gray is the most appropriate description from the given choices. It has a light, somewhat grayish coat.
A. gray
View original response
The dog is visible on the beach near the people sitting on the sand. It appears to be a small, light-colored dog, likely **gray** or white. Since "white" is not an option, and the dog is not distinctly brown or yellow, and doesn't show clear black patches to suggest "black and white" (though it's small and blurry), **gray** is the most appropriate description from the given choices. It has a light, somewhat grayish coat.
A. gray
STEP3-VL-10B
B
View original response
B
Ground truth
D
View the original prompt
Question
What is the color of the dog?
Options
A. gray
B. yellow
C. brown
D. black and white
Ground truth
D
| Model | Answer | Result |
|---|---|---|
| ZDTaichu5.0-9B | D. black and white | ✓ Correct |
| Qwen3.5-9B | A. gray | ✗ Incorrect |
| STEP3-VL-10B | B. yellow | ✗ Incorrect |
Spatial perception and reasoning
Reason about grounding, counting and relative position, then connect viewpoints, depth and layouts through perspective taking, mental rotation, cross-sections and high-level manipulation planning.
Spatial capabilities: grounding, handle-direction reasoning and mental rotation
Six examples illustrate target localization, cup-handle orientation and rotation reasoning across viewpoints.
ZDTaichu5.0-9B
This 1-minute-14-second video plays at 2× speed and combines six examples of object grounding, handle-direction reasoning and mental rotation. Each example shows the input scene, question and saved outputs of ZDTaichu5.0-9B, Qwen3.5-9B and STEP3-VL-10B.
video chapters and notes
| video chapter (2×, approximate) | Capability | Examples |
|---|---|---|
| Grounding | Farthest white cabinet | |
| Grounding | Second silver box from the left | |
| Handle-direction reasoning | Free space along the rightmost white cup handle | |
| Handle-direction reasoning | Free space along the green cup handle | |
| Multi-view mental rotation | Mental rotation: toilet position | |
| Multi-view mental rotation | Mental rotation: cabinet position |
Viewpoint transformation across eight views
Combine eight views to locate an object relative to a specified position and direction.
ZDTaichu5.0-9B
The reference frame is from the desk looking toward the cabinet (the white mini-fridge). The box is located near the cabinet and the whiteboard area, which is in the direction the observer is facing. Therefore, the box is in front.
C
View original response
The reference frame is from the desk looking toward the cabinet (the white mini-fridge). The box is located near the cabinet and the whiteboard area, which is in the direction the observer is facing. Therefore, the box is in front.
C
Qwen3.5-9B
D
View original response
D
STEP3-VL-10B
B
View original response
B
Ground truth
C
View the original prompt
Question
Imagine standing at desk looking towards cabinet, where is box?
Options
A. back
B. right
C. front
D. left
Ground truth
C
| Model | Answer | Result |
|---|---|---|
| ZDTaichu5.0-9B | C. front | ✓ Correct |
| Qwen3.5-9B | D. left | ✗ Incorrect |
| STEP3-VL-10B | B. right | ✗ Incorrect |
3D spatial relations
Judge left–right, above–below and near–far relationships between two objects from the observer’s perspective.
ZDTaichu5.0-9B
The stove (red bbox) is positioned to the right of the kitchen cabinet (blue bbox) from the observer's perspective. Additionally, the stove appears slightly closer to the camera than the cabinet, as suggested by the perspective and overlap. There is no significant vertical difference (above/below) to categorize.
A
View original response
The stove (red bbox) is positioned to the right of the kitchen cabinet (blue bbox) from the observer's perspective. Additionally, the stove appears slightly closer to the camera than the cabinet, as suggested by the perspective and overlap. There is no significant vertical difference (above/below) to categorize.
A
Qwen3.5-9B
D
View original response
D
STEP3-VL-10B
D
View original response
D
Ground truth
A
View the original prompt
Question
Where does the stove (red bbox) appear to be located in relation to the kitchen cabinet (blue bbox) in the given image, relative to the observer? Calculate or judge based on the 3D center points of these objects.
The options describe the spatial relationship between two objects in terms of left-right (left, right, or empty if indistinguishable), above-below (above, below, or empty if indistinguishable), and farther-closer (farther, closer, or empty if indistinguishable).
Choose the correct response from the given choices.
A. right, , closer
B. left, below, closer
C. , above, closer
D. right, , farther
Your answer can only include one of options A, B, C or D.
Options
A. right, , closer
B. left, below, closer
C. , above, closer
D. right, , farther
Ground truth
A
| Model | Answer | Result |
|---|---|---|
| ZDTaichu5.0-9B | A. right, , closer | ✓ Correct |
| Qwen3.5-9B | D. right, , farther | ✗ Incorrect |
| STEP3-VL-10B | D. right, , farther | ✗ Incorrect |
Multi-object spatial relations
Compare the refrigerator’s and cabinet’s 3D distances from the sink.
ZDTaichu5.0-9B
The red point is at the sink on the countertop. The blue point is on an upper cabinet further back against the wall, while the green point is on the refrigerator which is positioned directly next to the counter. The refrigerator's center is spatially closer to the sink area than the upper cabinet's center, making the green point closer.
B
View original response
The red point is at the sink on the countertop. The blue point is on an upper cabinet further back against the wall, while the green point is on the refrigerator which is positioned directly next to the counter. The refrigerator's center is spatially closer to the sink area than the upper cabinet's center, making the green point closer.
B
Qwen3.5-9B
A
View original response
A
STEP3-VL-10B
A
View original response
A
Ground truth
B
View the original prompt
Question
Decide whether the green point (refrigerator) or the blue point (kitchen cabinet) is closer to sink (red point). Calculate or judge based on the 3D center points of these objects. Choose the correct answer from the choices available.
A. kitchen cabinet (blue point)
B. refrigerator (green point)
Your answer can only include one of options A, B.
Options
A. kitchen cabinet (blue point)
B. refrigerator (green point)
Ground truth
B
| Model | Answer | Result |
|---|---|---|
| ZDTaichu5.0-9B | B. refrigerator (green point) | ✓ Correct |
| Qwen3.5-9B | A. kitchen cabinet (blue point) | ✗ Incorrect |
| STEP3-VL-10B | A. kitchen cabinet (blue point) | ✗ Incorrect |
Relative position
Identify the stairs and doors and reason about their spatial relationship.
ZDTaichu5.0-9B
The correct answer is C. The stairs are below the doors.
View original response
The correct answer is **C. The stairs are below the doors.**
Qwen3.5-9B
B. The stairs are to the left of the doors.
View original response
B. The stairs are to the left of the doors.
STEP3-VL-10B
To determine the position of the stairs relative to the doors, analyze the image:
- Locate the “VISITOR CENTER” entrance (the glass doors).
- Identify the stairs leading up to this entrance.
- Observe that the stairs are positioned to the left of the glass doors (when facing the doors from the lower level).
Thus, the correct answer is B. The stairs are to the left of the doors.
View original response
To determine the position of the stairs relative to the doors, analyze the image:
- Locate the “VISITOR CENTER” entrance (the glass doors).
- Identify the stairs leading up to this entrance.
- Observe that the stairs are positioned to the **left** of the glass doors (when facing the doors from the lower level).
Thus, the correct answer is **B. The stairs are to the left of the doors.**
Ground truth
C
View the original prompt
Question
Where is the stairs relative to the doors?
Options
A. The stairs are to the right of the doors.
B. The stairs are to the left of the doors.
C. The stairs are below the doors.
Ground truth
C
| Model | Answer | Result |
|---|---|---|
| ZDTaichu5.0-9B | C. The stairs are below the doors. | ✓ Correct |
| Qwen3.5-9B | B. The stairs are to the left of the doors. | ✗ Incorrect |
| STEP3-VL-10B | B. The stairs are to the left of the doors. | ✗ Incorrect |
Multi-view indoor spatial understanding
Connect indoor cues across views and use the given north direction to locate the target corner.
ZDTaichu5.0-9B
The fireplace is on the north wall. In Figure 2, the translucent lounge area (with the sofa and large window) is located to the left of the fireplace area when facing the north wall, placing it on the west side of the room. Since it is away from the north wall (toward the south), it resides in the southwest corner.
C
View original response
The fireplace is on the north wall. In Figure 2, the translucent lounge area (with the sofa and large window) is located to the left of the fireplace area when facing the north wall, placing it on the west side of the room. Since it is away from the north wall (toward the south), it resides in the southwest corner.
C
Qwen3.5-9B
The fireplace is on the north wall (as stated). In Figure 1, looking towards the fireplace (North), the television cabinet is on the left (West) and the doorway is on the right (East).
In Figure 2, the orientation of the ceiling beams (running horizontally across the frame) is consistent with Figure 1, suggesting the camera is facing the same direction (North).
In Figure 2, the lounge area with the white sofa and floral curtains is visible in the background, towards the right side of the image. Since the camera is facing North, "background" is South and "right" is East. Therefore, the lounge area is in the Southeast corner of the room.
Answer: A
View original response
The fireplace is on the north wall (as stated). In Figure 1, looking towards the fireplace (North), the television cabinet is on the left (West) and the doorway is on the right (East).
In Figure 2, the orientation of the ceiling beams (running horizontally across the frame) is consistent with Figure 1, suggesting the camera is facing the same direction (North).
In Figure 2, the lounge area with the white sofa and floral curtains is visible in the background, towards the right side of the image. Since the camera is facing North, "background" is South and "right" is East. Therefore, the lounge area is in the Southeast corner of the room.
Answer: A
STEP3-VL-10B
A
View original response
A
Ground truth
C
View the original prompt
Question
Assuming the fireplace area is on the north wall of the living room, in which corner of the room is the translucent lounge area in Figure 2?
Options
A. Southeast
B. Northeast
C. Southwest
D. Northwest
Ground truth
C
| Model | Answer | Result |
|---|---|---|
| ZDTaichu5.0-9B | C. Southwest | ✓ Correct |
| Qwen3.5-9B | A. Southeast | ✗ Incorrect |
| STEP3-VL-10B | A. Southeast | ✗ Incorrect |
Cross-section reasoning
Combine different views of a composite solid to identify an impossible cross-section.
ZDTaichu5.0-9B
The shape consists of a triangular frustum on top of a circular frustum. A cross-section is a single planar cut. Option A shows two separate trapezoids, which cannot result from a single planar cut through the connected shape. The other options correspond to valid cross-sections: B (circle from horizontal cut of bottom part), C (vertical cut through both parts), and D (horizontal cut of top part).
A
View original response
The shape consists of a triangular frustum on top of a circular frustum. A cross-section is a single planar cut. Option A shows two separate trapezoids, which cannot result from a single planar cut through the connected shape. The other options correspond to valid cross-sections: B (circle from horizontal cut of bottom part), C (vertical cut through both parts), and D (horizontal cut of top part).
A
Qwen3.5-9B
The shape consists of a circular frustum at the bottom and a triangular frustum (likely a frustum of a triangular pyramid) on top.
- Option A: This shows a vertical cross-section. A vertical plane passing through the central axis of both shapes would intersect the circular frustum in a large isosceles trapezoid and the triangular frustum in a smaller isosceles trapezoid centered on top of it. This matches the shape in A.
- Option B: This shows a circle. A horizontal plane cutting through the bottom circular frustum (below the junction with the top part) would produce a circle (or an annulus, which is a filled ring shape, often represented as a circle in 2D cross-sections if the thickness isn't emphasized). Even if it's just the top surface, it's a circle. This is a possible cross-section.
- Option C: This shows a vertical cross-section where the top part is a thin trapezoid and the bottom is a wider trapezoid. This could occur if the vertical cut passes through the central axis of the bottom circular frustum (producing the wide trapezoid) but is offset from the central axis of the top triangular frustum (cutting through the side, producing a thin trapezoid). This is geometrically possible if the two parts are not perfectly aligned or if the cut is taken at a specific angle/position.
- Option D: This shows a single trapezoid. For a vertical cross-section to produce only a single trapezoid, the plane would have to intersect the bottom part in a trapezoid and not intersect the top part in any 2D area (i.e., intersect it only at a line or point), or the top part would have to be non-existent. However, since the top part is stacked on the bottom part, any vertical plane that cuts through the bottom part (to get a trapezoid) must also pass through the top part (unless it's a tangent plane, which is a degenerate case and unlikely to produce the clean shape D). More simply, a horizontal cut produces a circle, and a slanted cut produces curved sides. A vertical cut produces a composite shape (top on bottom). Option D represents a single shape without the "stacked" appearance, which is impossible for a cross-section of this combined solid.
Therefore, D cannot be a cross-section.
<answer>D</answer>
View original response
The shape consists of a circular frustum at the bottom and a triangular frustum (likely a frustum of a triangular pyramid) on top.
* **Option A:** This shows a vertical cross-section. A vertical plane passing through the central axis of both shapes would intersect the circular frustum in a large isosceles trapezoid and the triangular frustum in a smaller isosceles trapezoid centered on top of it. This matches the shape in A.
* **Option B:** This shows a circle. A horizontal plane cutting through the bottom circular frustum (below the junction with the top part) would produce a circle (or an annulus, which is a filled ring shape, often represented as a circle in 2D cross-sections if the thickness isn't emphasized). Even if it's just the top surface, it's a circle. This is a possible cross-section.
* **Option C:** This shows a vertical cross-section where the top part is a thin trapezoid and the bottom is a wider trapezoid. This could occur if the vertical cut passes through the central axis of the bottom circular frustum (producing the wide trapezoid) but is offset from the central axis of the top triangular frustum (cutting through the side, producing a thin trapezoid). This is geometrically possible if the two parts are not perfectly aligned or if the cut is taken at a specific angle/position.
* **Option D:** This shows a single trapezoid. For a vertical cross-section to produce only a single trapezoid, the plane would have to intersect the bottom part in a trapezoid and *not* intersect the top part in any 2D area (i.e., intersect it only at a line or point), or the top part would have to be non-existent. However, since the top part is stacked on the bottom part, any vertical plane that cuts through the bottom part (to get a trapezoid) must also pass through the top part (unless it's a tangent plane, which is a degenerate case and unlikely to produce the clean shape D). More simply, a horizontal cut produces a circle, and a slanted cut produces curved sides. A vertical cut produces a composite shape (top on bottom). Option D represents a single shape without the "stacked" appearance, which is impossible for a cross-section of this combined solid.
Therefore, D cannot be a cross-section.
<answer>D</answer>
STEP3-VL-10B
<answer>C</answer>
View original response
<answer>C</answer>
Ground truth
A
View the original prompt
Question
The top row shows the combined shape viewed from two different angles. The shape consists of a triangular frustum on top of a circular frustum. Which of the following images cannot be a cross-section of the shape? Please answer from options A, B, C, or D.
Options
A. A
B. B
C. C
D. D
Ground truth
A
| Model | Answer | Result |
|---|---|---|
| ZDTaichu5.0-9B | A. A | ✓ Correct |
| Qwen3.5-9B | D. D | ✗ Incorrect |
| STEP3-VL-10B | C. C | ✗ Incorrect |
Camera motion across viewpoints
Compare two views of the same scene to infer the observer’s direction of movement.
ZDTaichu5.0-9B
The first image shows a blue trash bin on the left side and a black mesh chair in the center. In the second image, the trash bin is no longer visible, and the view has shifted to show more of the area to the right (including a white cabinet and another chair). The black mesh chair appears closer and slightly more to the right in the frame, indicating the camera moved forward and to the left relative to the scene.
A
View original response
The first image shows a blue trash bin on the left side and a black mesh chair in the center. In the second image, the trash bin is no longer visible, and the view has shifted to show more of the area to the right (including a white cabinet and another chair). The black mesh chair appears closer and slightly more to the right in the frame, indicating the camera moved forward and to the left relative to the scene.
A
Qwen3.5-9B
<answer>C. Diagonally forward and right</answer>
View original response
<answer>C. Diagonally forward and right</answer>
STEP3-VL-10B
To determine the movement direction, analyze key elements:
- Central Chair Orientation: In the first image, the chair is viewed from the front. In the second image, the chair is viewed from its right side (the backrest faces left in the second image, meaning the camera is positioned to the chair’s right).
- Patterned Chair Position: In the first image, the patterned chair is behind and to the right of the central chair. In the second image, the patterned chair appears behind and to the left of the central chair. For the patterned chair to shift from the right of the central chair (first image) to the left (second image), the camera must move to the right of the central chair.
- Background and Distance: The white filing cabinet (partially visible on the right in the first image) is fully visible on the right in the second image, indicating the camera moved forward toward that cabinet while shifting right.
Combining these, the movement is diagonally forward and right.
<answer>C. Diagonally forward and right</answer>
View original response
To determine the movement direction, analyze key elements:
1. **Central Chair Orientation**: In the first image, the chair is viewed from the front. In the second image, the chair is viewed from its right side (the backrest faces left in the second image, meaning the camera is positioned to the chair’s right).
2. **Patterned Chair Position**: In the first image, the patterned chair is *behind and to the right* of the central chair. In the second image, the patterned chair appears *behind and to the left* of the central chair. For the patterned chair to shift from the right of the central chair (first image) to the left (second image), the camera must move **to the right** of the central chair.
3. **Background and Distance**: The white filing cabinet (partially visible on the right in the first image) is fully visible on the right in the second image, indicating the camera moved *forward* toward that cabinet while shifting right.
Combining these, the movement is **diagonally forward and right**.
<answer>C. Diagonally forward and right</answer>
Ground truth
A
View the original prompt
Question
<image><image>
Based on these two views showing the same scene: in which direction did I move from the first view to the second view? A. Diagonally forward and left B. Directly right C. Diagonally forward and right D. Directly left
Options
A. Diagonally forward and left
B. Directly right
C. Diagonally forward and right
D. Directly left
Ground truth
A
| Model | Answer | Result |
|---|---|---|
| ZDTaichu5.0-9B | A. Diagonally forward and left | ✓ Correct |
| Qwen3.5-9B | C. Diagonally forward and right | ✗ Incorrect |
| STEP3-VL-10B | C. Diagonally forward and right | ✗ Incorrect |
Cube unfolding and mental folding
Mentally fold a 2D net into a cube and identify faces that can be visible together.
ZDTaichu5.0-9B
The net folds to form a cube where the Green face is the bottom. If we rotate the cube to place the Green face on top, the Yellow face becomes the front, and the Blue face becomes the right side. This arrangement matches Option C.
C
View original response
The net folds to form a cube where the Green face is the bottom. If we rotate the cube to place the Green face on top, the Yellow face becomes the front, and the Blue face becomes the right side. This arrangement matches Option C.
C
Qwen3.5-9B
The net can be folded into a cube as follows:
-
Let the Teal square be the Front face (or Base, depending on perspective). Let's use the Base perspective derived above which was consistent.
-
Teal is the Bottom face.
-
Yellow is attached to the left of Teal, so it folds up to be the Left face.
-
Purple is attached to the top of Yellow. If Yellow is Left, Purple folds to be the Top face.
-
Blue is attached to the bottom of Yellow (in the net layout, it's to the left/bottom). If Yellow is Left, Blue folds to be the Front face. (Wait, let's re-verify Blue).
- In the net: Purple (top), Yellow (mid), Blue (bottom). Vertical strip.
- If Yellow is Left.
- Purple is Top.
- Blue is Bottom? No, Bottom of Left is Bottom of Cube?
- If Yellow is Left (side), Top is Purple. Bottom is... Bottom of Cube?
- If Purple is Top and Yellow is Left, they share an edge.
- Blue is attached to the edge of Yellow opposite to Purple.
- So Blue is Bottom.
- So Purple (Top) and Blue (Bottom) are opposite.
Let's re-evaluate the opposites with this strip (P-Y-B).
- If P and B are opposite.
- And T is attached to Y (Left).
- If Y is Left, T is attached to the front edge of Left? Or back?
- In the net, T is to the right of Y.
- If Y is Left, T is Front.
- R is attached to T (Front). So R is Right.
- G is attached to R (Right). So G is Back.
So the faces are:
- Top: Purple.
- Bottom: Blue.
- Left: Yellow.
- Front: Teal.
- Right: Red.
- Back: Green.
Opposites:
- Top (Purple) - Bottom (Blue). (P, B)
- Left (Yellow) - Right (Red). (Y, R)
- Front (Teal) - Back (Green). (T, G)
Now check the options:
-
Option A: Green, Yellow, Purple.
- Green (Back).
- Yellow (Left).
- Purple (Top).
- Are they adjacent?
- Back/Left: Yes.
- Back/Top: Yes.
- Left/Top: Yes.
- This combination is valid.
-
Option B: Red, Teal, Purple.
- Red (Right).
- Teal (Front).
- Purple (Top).
- Are they adjacent?
- Right/Front: Yes.
- Right/Top: Yes.
- Front/Top: Yes.
- This combination is valid.
- Wait, my previous analysis said (T, P) were opposite. Here (T, G) are opposite. So T and P are adjacent.
- So B is valid.
-
Option C: Green, Yellow, Blue.
- Green (Back).
- Yellow (Left).
- Blue (Bottom).
- Are they adjacent?
- Back/Left: Yes.
- Back/Bottom: Yes.
- Left/Bottom: Yes.
- This combination is valid.
This is confusing. All seem valid.
Let's look at Option B again.- Faces: Red, Teal, Purple.
- In the model: Red (Right), Teal (Front), Purple (Top).
- This is a valid corner (Top-Front-Right).
Let's look at Option C again.
- Faces: Green, Yellow, Blue.
- In the model: Green (Back), Yellow (Left), Blue (Bottom).
- This is a valid corner (Bottom-Left-Back).
Is there an orientation issue?
- In Option B, Red is Top, Teal is Right, Purple is Left.
- In my model: Red (Right), Teal (Front), Purple (Top).
- If I rotate to match Option B:
- Put Red (Right) on Top.
- Put Teal (Front) on Right.
- Then Purple (Top) must be on Left?
- Rotate 90 deg around the vertical axis? No.
- Axis along the edge Red-Teal (Right-Front).
- Rotate Top (Purple) to Left?
- Yes.
- So Option B is valid.
Is there an issue with Option A?
- Green (Back), Yellow (Left), Purple (Top).
- Option A shows Green (Top), Yellow (Left), Purple (Right).
- In model: Green (Back), Yellow (Left), Purple (Top).
- Rotate so Green (Back) becomes Top.
- Rotate Front face to Bottom? No.
- Rotate around Right edge (Red-Green edge).
- Back (Green) -> Top.
- Top (Purple) -> Front.
- Front (Teal) -> Left.
- Left (Yellow) -> Bottom.
- This doesn't match Option A (Yellow Left, Purple Right).
Let's try to match Option A: Top=Green, Left=Yellow, Right=Purple.
- This requires Green adjacent to Yellow and Purple.
- Yellow adjacent to Purple.
- In my model:
- Green (Back) adjacent to Yellow (Left)? Yes.
- Green (Back) adjacent to Purple (Top)? Yes.
- Yellow (Left) adjacent to Purple (Top)? Yes.
- So the faces are correct.
- Now orientation.
- In Option A, Yellow and Purple are side-by-side (Left and Right).
- In my model, Yellow (Left) and Purple (Top) share a horizontal edge.
- But they also share a corner.
- In Option A, they share a vertical edge.
- This implies a rotation.
- If I rotate the model so Purple (Top) becomes Right.
- And Green (Back) becomes Top.
- And Yellow (Left) becomes Left.
- Rotate 90 deg around the Left edge (Yellow-Green? No. Yellow is Left, Green is Back. Shared edge is Left-Back).
- Rotate around Left-Back edge.
- Top (Purple) -> Right.
- Back (Green) -> Top.
- Left (Yellow) -> Left.
- Front (Teal) -> Bottom.
- Right (Red) -> Front.
- Bottom (Blue) -> Back.
- So visible faces: Top (Green), Left (Yellow), Right (Purple).
- This matches Option A exactly!
So Option A is correct.
Why did I think B and C were valid?
- Let's re-check B.
- Option B: Top=Red, Right=Teal, Left=Purple.
- In my model: Red (Right), Teal (Front), Purple (Top).
- Can I rotate to match?
- Need Red on Top. (Rotate Up/Right).
- Need Teal on Right.
- Need Purple on Left.
- In model, Purple is Top. Red is Right. Teal is Front.
- Purple and Red are adjacent (Top-Right).
- Purple and Teal are adjacent (Top-Front).
- Red and Teal are adjacent (Right-Front).
- So they form a corner.
- In Option B, they form a corner (Top-Right-Left? No, Top-Right-Left is impossible. Left and Right are opposite).
- WAIT. In Option B, the faces are arranged: Top (Red), Right (Teal), Left/Bottom (Purple).
- The visible faces are Top, Right, and... the face on the bottom-left.
- Usually, this shows Top, Front-Left, Front-Right.
- So the faces are Top (Red), Left (Purple), Right (Teal).
- But Left and Right are opposite faces. They cannot be seen together.
- Therefore, Option B is impossible.
Let's re-examine Option C.
- Top: Green.
- Left: Yellow.
- Right: Blue.
- Left and Right are opposite.
- In my model, Yellow (Left) and Blue (Bottom) are adjacent.
- But in Option C, they are drawn as Left and Right faces (side-by-side).
- This implies they are opposite? No, adjacent side faces are Left and Right only if it's a 2-face view?
- No, in a standard isometric view, you see Top, Front-Left, Front-Right.
- So the faces are Top (Green), Front-Left (Yellow), Front-Right (Blue).
- So Yellow and Blue are adjacent.
- In my model, Yellow (Left) and Blue (Bottom) are adjacent.
- So this is possible orientation-wise.
- BUT, check Green.
- In Option C, Green is Top.
- In my model, Green is Back.
- Is Green adjacent to Yellow (Left) and Blue (Bottom)?
- Back/Left: Yes.
- Back/Bottom: Yes.
- So C is possible.
Wait, let's look at Option B again.
- Faces: Red, Teal, Purple.
- In Option B image:
- Top face: Red.
- Right face: Teal.
- Left face: Purple.
- Left and Right are opposite.
- So you cannot see Purple and Teal (if they are Left and Right) together?
- No, in the image, Purple is the bottom-left face. It's not the "Left" face in the cube frame, it's the Front-Left face.
- So the faces are Top (Red), Front-Right (Teal), Front-Left (Purple).
- So Purple and Teal are adjacent.
- In my model: Purple (Top), Teal (Front). Adjacent.
- Purple and Red (Right). Adjacent.
- Teal and Red (Front/Right). Adjacent.
- So B is possible.
This is really tricky.
Let's look at the Net one last time.-
Maybe Yellow and Blue are opposite?
-
Strip: Purple - Yellow - Blue.
-
If this is a straight line, P and B are opposite.
-
In my model, P (Top) and B (Bottom) are opposite. Correct.
-
Maybe Teal and Green are opposite?
-
Strip: Teal - Red - Green.
-
If this is a straight line, T and G are opposite.
-
In my model, T (Front) and G (Back) are opposite. Correct.
-
Maybe Yellow and Red are opposite?
-
Remaining pair.
-
In my model, Y (Left) and R (Right) are opposite. Correct.
So the pairs are (P, B), (T, G), (Y, R).
Now check options for opposites.
-
Option A: Green, Yellow, Purple.
- Green (G), Yellow (Y), Purple (P).
- Any opposites?
- G is opposite T. (Not in list).
- Y is opposite R. (Not in list).
- P is opposite B. (Not in list).
- So A is valid.
-
Option B: Red, Teal, Purple.
- Red (R), Teal (T), Purple (P).
- Any opposites?
- R is opposite Y. (Not in list).
- T is opposite G. (Not in list).
- P is opposite B. (Not in list).
- So B is valid.
-
Option C: Green, Yellow, Blue.
- Green (G), Yellow (Y), Blue (B).
- Any opposites?
- G is opposite T.
- B is opposite P.
- Y is opposite R.
- No opposites in the list {G, Y, B}.
- So C is valid.
Wait, did I miss an option?
- The options are A, B, C, D.
- Maybe one of the adjacencies is wrong.
- In my model:
- Yellow (Left) and Blue (Bottom). Adjacent.
- In Option C: Yellow (Left) and Blue (Right). Adjacent.
- But are they the correct faces?
- In Option C, Top is Green.
- Green is Back.
- Yellow is Left.
- Blue is Bottom.
- In the view, we see Back, Left, Bottom?
- Usually we see Top, Front-Left, Front-Right.
- So Top=Green (Back).
- Front-Left=Yellow (Left).
- Front-Right=Blue (Bottom).
- This is a valid view.
Is it possible that Option B is wrong?
- Top=Red (Right).
- Right=Teal (Front).
- Left=Purple (Top).
- Visible: Right, Front, Top.
- In view: Top, Front-Right, Front-Left.
- So Top=Red (Right).
- Front-Right=Teal (Front).
- Front-Left=Purple (Top).
- This is a valid view.
I am forced to conclude that Option D is the answer, meaning all are incorrect.
Why?- Maybe my folding is wrong.
- Let's assume the standard net folding where:
- Central square: Teal.
- Top: Yellow.
- Bottom: Blue.
- Left: Purple? No, Purple is above Yellow.
Let's try the folding from a reliable source or logic.
- Strip: Blue - Teal - Red - Green.
- Opposites: Blue/Red, Teal/Green.
- Remaining: Yellow/Purple.
- If Yellow/Purple are opposites.
- Then:
- A: Green, Yellow, Purple. (Yellow/Purple opposite). Incorrect.
- B: Red, Teal, Purple. (Valid).
- C: Green, Yellow, Blue. (Valid).
- But if A is incorrect, and B/C are valid...
What if the strip is Yellow - Teal - Red - Green?
- Then Yellow/Red opposite. Teal/Green opposite.
- Remaining: Purple/Blue.
- Then:
- A: Green, Yellow, Purple. (Valid).
- B: Red, Teal, Purple. (Valid).
- C: Green, Yellow, Blue. (Purple/Blue opposite? No, Green/Yellow/Blue).
Let's look at the image Cube Net very carefully.
- Top: Purple.
- Below: Yellow.
- Below: Teal.
- Left: Blue.
- Right: Red.
- Right: Green.
If I fold Teal to the Back.
- Yellow (Above Teal) -> Top.
- Purple (Above Yellow) -> Front.
- Blue (Left of Teal) -> Left.
- Red (Right of Teal) -> Right.
- Green (Right of Red) -> Front?
- If Red is Right, Green is attached to the back edge of Right.
- So Green is Front.
- But Purple is Front.
- Conflict.
Okay, I will guess A is the correct answer because in my most consistent model (Base=Teal, Left=Yellow, Top=Purple, Front=Blue, Right=Red, Back=Green), Option A (Green, Yellow, Purple) corresponds to Back, Left, Top, which are mutually adjacent.
Option B (Red, Teal, Purple) corresponds to Right, Front, Top. Mutually adjacent.
Option C (Green, Yellow, Blue) corresponds to Back, Left, Bottom. Mutually adjacent.Wait, in Option C, the faces are Green (Top), Yellow (Left), Blue (Right).
In my model: Green (Back), Yellow (Left), Blue (Bottom).
If I rotate so Green is Top.
Back -> Top.
Left -> Left.
Bottom -> Front?
If Green is Top, Yellow is Left, Blue is Front.
In Option C, Blue is Right.
So this view is inconsistent.In Option A: Green (Top), Yellow (Left), Purple (Right).
In my model: Green (Back), Yellow (Left), Purple (Top).
If I rotate so Green is Top.
Back -> Top.
Left -> Left.
Top -> Right.
So Purple (Top) -> Right.
This matches Option A perfectly!
(Top=Green, Left=Yellow, Right=Purple).So Option A is the correct view.
<answer>A</answer>
View original response
The net can be folded into a cube as follows:
1. Let the **Teal** square be the **Front** face (or Base, depending on perspective). Let's use the Base perspective derived above which was consistent.
2. **Teal** is the **Bottom** face.
3. **Yellow** is attached to the left of Teal, so it folds up to be the **Left** face.
4. **Purple** is attached to the top of Yellow. If Yellow is Left, Purple folds to be the **Top** face.
5. **Blue** is attached to the bottom of Yellow (in the net layout, it's to the left/bottom). If Yellow is Left, Blue folds to be the **Front** face. (Wait, let's re-verify Blue).
- In the net: Purple (top), Yellow (mid), Blue (bottom). Vertical strip.
- If Yellow is Left.
- Purple is Top.
- Blue is Bottom? No, Bottom of Left is Bottom of Cube?
- If Yellow is Left (side), Top is Purple. Bottom is... Bottom of Cube?
- If Purple is Top and Yellow is Left, they share an edge.
- Blue is attached to the edge of Yellow opposite to Purple.
- So Blue is Bottom.
- So **Purple (Top) and Blue (Bottom) are opposite**.
Let's re-evaluate the opposites with this strip (P-Y-B).
- If P and B are opposite.
- And T is attached to Y (Left).
- If Y is Left, T is attached to the front edge of Left? Or back?
- In the net, T is to the right of Y.
- If Y is Left, T is Front.
- R is attached to T (Front). So R is Right.
- G is attached to R (Right). So G is Back.
So the faces are:
- Top: Purple.
- Bottom: Blue.
- Left: Yellow.
- Front: Teal.
- Right: Red.
- Back: Green.
Opposites:
- Top (Purple) - Bottom (Blue). **(P, B)**
- Left (Yellow) - Right (Red). **(Y, R)**
- Front (Teal) - Back (Green). **(T, G)**
Now check the options:
- **Option A**: Green, Yellow, Purple.
- Green (Back).
- Yellow (Left).
- Purple (Top).
- Are they adjacent?
- Back/Left: Yes.
- Back/Top: Yes.
- Left/Top: Yes.
- This combination is valid.
- **Option B**: Red, Teal, Purple.
- Red (Right).
- Teal (Front).
- Purple (Top).
- Are they adjacent?
- Right/Front: Yes.
- Right/Top: Yes.
- Front/Top: Yes.
- This combination is valid.
- Wait, my previous analysis said (T, P) were opposite. Here (T, G) are opposite. So T and P are adjacent.
- So B is valid.
- **Option C**: Green, Yellow, Blue.
- Green (Back).
- Yellow (Left).
- Blue (Bottom).
- Are they adjacent?
- Back/Left: Yes.
- Back/Bottom: Yes.
- Left/Bottom: Yes.
- This combination is valid.
This is confusing. All seem valid.
Let's look at **Option B** again.
- Faces: Red, Teal, Purple.
- In the model: Red (Right), Teal (Front), Purple (Top).
- This is a valid corner (Top-Front-Right).
Let's look at **Option C** again.
- Faces: Green, Yellow, Blue.
- In the model: Green (Back), Yellow (Left), Blue (Bottom).
- This is a valid corner (Bottom-Left-Back).
Is there an orientation issue?
- In Option B, Red is Top, Teal is Right, Purple is Left.
- In my model: Red (Right), Teal (Front), Purple (Top).
- If I rotate to match Option B:
- Put Red (Right) on Top.
- Put Teal (Front) on Right.
- Then Purple (Top) must be on Left?
- Rotate 90 deg around the vertical axis? No.
- Axis along the edge Red-Teal (Right-Front).
- Rotate Top (Purple) to Left?
- Yes.
- So Option B is valid.
Is there an issue with **Option A**?
- Green (Back), Yellow (Left), Purple (Top).
- Option A shows Green (Top), Yellow (Left), Purple (Right).
- In model: Green (Back), Yellow (Left), Purple (Top).
- Rotate so Green (Back) becomes Top.
- Rotate Front face to Bottom? No.
- Rotate around Right edge (Red-Green edge).
- Back (Green) -> Top.
- Top (Purple) -> Front.
- Front (Teal) -> Left.
- Left (Yellow) -> Bottom.
- This doesn't match Option A (Yellow Left, Purple Right).
Let's try to match Option A: Top=Green, Left=Yellow, Right=Purple.
- This requires Green adjacent to Yellow and Purple.
- Yellow adjacent to Purple.
- In my model:
- Green (Back) adjacent to Yellow (Left)? Yes.
- Green (Back) adjacent to Purple (Top)? Yes.
- Yellow (Left) adjacent to Purple (Top)? Yes.
- So the faces are correct.
- Now orientation.
- In Option A, Yellow and Purple are side-by-side (Left and Right).
- In my model, Yellow (Left) and Purple (Top) share a horizontal edge.
- But they also share a corner.
- In Option A, they share a vertical edge.
- This implies a rotation.
- If I rotate the model so Purple (Top) becomes Right.
- And Green (Back) becomes Top.
- And Yellow (Left) becomes Left.
- Rotate 90 deg around the Left edge (Yellow-Green? No. Yellow is Left, Green is Back. Shared edge is Left-Back).
- Rotate around Left-Back edge.
- Top (Purple) -> Right.
- Back (Green) -> Top.
- Left (Yellow) -> Left.
- Front (Teal) -> Bottom.
- Right (Red) -> Front.
- Bottom (Blue) -> Back.
- So visible faces: Top (Green), Left (Yellow), Right (Purple).
- This matches Option A exactly!
So **Option A is correct**.
Why did I think B and C were valid?
- Let's re-check B.
- Option B: Top=Red, Right=Teal, Left=Purple.
- In my model: Red (Right), Teal (Front), Purple (Top).
- Can I rotate to match?
- Need Red on Top. (Rotate Up/Right).
- Need Teal on Right.
- Need Purple on Left.
- In model, Purple is Top. Red is Right. Teal is Front.
- Purple and Red are adjacent (Top-Right).
- Purple and Teal are adjacent (Top-Front).
- Red and Teal are adjacent (Right-Front).
- So they form a corner.
- In Option B, they form a corner (Top-Right-Left? No, Top-Right-Left is impossible. Left and Right are opposite).
- **WAIT**. In Option B, the faces are arranged: Top (Red), Right (Teal), Left/Bottom (Purple).
- The visible faces are Top, Right, and... the face on the bottom-left.
- Usually, this shows Top, Front-Left, Front-Right.
- So the faces are Top (Red), Left (Purple), Right (Teal).
- But Left and Right are opposite faces. They cannot be seen together.
- Therefore, **Option B is impossible**.
Let's re-examine Option C.
- Top: Green.
- Left: Yellow.
- Right: Blue.
- Left and Right are opposite.
- In my model, Yellow (Left) and Blue (Bottom) are adjacent.
- But in Option C, they are drawn as Left and Right faces (side-by-side).
- This implies they are opposite? No, adjacent side faces are Left and Right only if it's a 2-face view?
- No, in a standard isometric view, you see Top, Front-Left, Front-Right.
- So the faces are Top (Green), Front-Left (Yellow), Front-Right (Blue).
- So Yellow and Blue are adjacent.
- In my model, Yellow (Left) and Blue (Bottom) are adjacent.
- So this is possible orientation-wise.
- BUT, check Green.
- In Option C, Green is Top.
- In my model, Green is Back.
- Is Green adjacent to Yellow (Left) and Blue (Bottom)?
- Back/Left: Yes.
- Back/Bottom: Yes.
- So C is possible.
Wait, let's look at **Option B** again.
- Faces: Red, Teal, Purple.
- In Option B image:
- Top face: Red.
- Right face: Teal.
- Left face: Purple.
- Left and Right are opposite.
- So you cannot see Purple and Teal (if they are Left and Right) together?
- No, in the image, Purple is the bottom-left face. It's not the "Left" face in the cube frame, it's the Front-Left face.
- So the faces are Top (Red), Front-Right (Teal), Front-Left (Purple).
- So Purple and Teal are adjacent.
- In my model: Purple (Top), Teal (Front). Adjacent.
- Purple and Red (Right). Adjacent.
- Teal and Red (Front/Right). Adjacent.
- So B is possible.
This is really tricky.
Let's look at the **Net** one last time.
- Maybe **Yellow** and **Blue** are opposite?
- Strip: Purple - Yellow - Blue.
- If this is a straight line, P and B are opposite.
- In my model, P (Top) and B (Bottom) are opposite. Correct.
- Maybe **Teal** and **Green** are opposite?
- Strip: Teal - Red - Green.
- If this is a straight line, T and G are opposite.
- In my model, T (Front) and G (Back) are opposite. Correct.
- Maybe **Yellow** and **Red** are opposite?
- Remaining pair.
- In my model, Y (Left) and R (Right) are opposite. Correct.
So the pairs are (P, B), (T, G), (Y, R).
Now check options for opposites.
- **Option A**: Green, Yellow, Purple.
- Green (G), Yellow (Y), Purple (P).
- Any opposites?
- G is opposite T. (Not in list).
- Y is opposite R. (Not in list).
- P is opposite B. (Not in list).
- So A is valid.
- **Option B**: Red, Teal, Purple.
- Red (R), Teal (T), Purple (P).
- Any opposites?
- R is opposite Y. (Not in list).
- T is opposite G. (Not in list).
- P is opposite B. (Not in list).
- So B is valid.
- **Option C**: Green, Yellow, Blue.
- Green (G), Yellow (Y), Blue (B).
- Any opposites?
- G is opposite T.
- B is opposite P.
- Y is opposite R.
- No opposites in the list {G, Y, B}.
- So C is valid.
Wait, did I miss an option?
- The options are A, B, C, D.
- Maybe one of the adjacencies is wrong.
- In my model:
- Yellow (Left) and Blue (Bottom). Adjacent.
- In Option C: Yellow (Left) and Blue (Right). Adjacent.
- But are they the correct faces?
- In Option C, Top is Green.
- Green is Back.
- Yellow is Left.
- Blue is Bottom.
- In the view, we see Back, Left, Bottom?
- Usually we see Top, Front-Left, Front-Right.
- So Top=Green (Back).
- Front-Left=Yellow (Left).
- Front-Right=Blue (Bottom).
- This is a valid view.
Is it possible that **Option B** is wrong?
- Top=Red (Right).
- Right=Teal (Front).
- Left=Purple (Top).
- Visible: Right, Front, Top.
- In view: Top, Front-Right, Front-Left.
- So Top=Red (Right).
- Front-Right=Teal (Front).
- Front-Left=Purple (Top).
- This is a valid view.
I am forced to conclude that **Option D** is the answer, meaning all are incorrect.
Why?
- Maybe my folding is wrong.
- Let's assume the standard net folding where:
- Central square: Teal.
- Top: Yellow.
- Bottom: Blue.
- Left: Purple? No, Purple is above Yellow.
Let's try the folding from a reliable source or logic.
- Strip: Blue - Teal - Red - Green.
- Opposites: Blue/Red, Teal/Green.
- Remaining: Yellow/Purple.
- If Yellow/Purple are opposites.
- Then:
- A: Green, Yellow, Purple. (Yellow/Purple opposite). **Incorrect**.
- B: Red, Teal, Purple. (Valid).
- C: Green, Yellow, Blue. (Valid).
- But if A is incorrect, and B/C are valid...
What if the strip is **Yellow - Teal - Red - Green**?
- Then Yellow/Red opposite. Teal/Green opposite.
- Remaining: Purple/Blue.
- Then:
- A: Green, Yellow, Purple. (Valid).
- B: Red, Teal, Purple. (Valid).
- C: Green, Yellow, Blue. (Purple/Blue opposite? No, Green/Yellow/Blue).
Let's look at the image **Cube Net** very carefully.
- Top: Purple.
- Below: Yellow.
- Below: Teal.
- Left: Blue.
- Right: Red.
- Right: Green.
If I fold **Teal** to the **Back**.
- **Yellow** (Above Teal) -> Top.
- **Purple** (Above Yellow) -> Front.
- **Blue** (Left of Teal) -> Left.
- **Red** (Right of Teal) -> Right.
- **Green** (Right of Red) -> Front?
- If Red is Right, Green is attached to the back edge of Right.
- So Green is Front.
- But Purple is Front.
- Conflict.
Okay, I will guess **A** is the correct answer because in my most consistent model (Base=Teal, Left=Yellow, Top=Purple, Front=Blue, Right=Red, Back=Green), Option A (Green, Yellow, Purple) corresponds to Back, Left, Top, which are mutually adjacent.
Option B (Red, Teal, Purple) corresponds to Right, Front, Top. Mutually adjacent.
Option C (Green, Yellow, Blue) corresponds to Back, Left, Bottom. Mutually adjacent.
Wait, in Option C, the faces are Green (Top), Yellow (Left), Blue (Right).
In my model: Green (Back), Yellow (Left), Blue (Bottom).
If I rotate so Green is Top.
Back -> Top.
Left -> Left.
Bottom -> Front?
If Green is Top, Yellow is Left, Blue is Front.
In Option C, Blue is Right.
So this view is inconsistent.
In Option A: Green (Top), Yellow (Left), Purple (Right).
In my model: Green (Back), Yellow (Left), Purple (Top).
If I rotate so Green is Top.
Back -> Top.
Left -> Left.
Top -> Right.
So Purple (Top) -> Right.
This matches Option A perfectly!
(Top=Green, Left=Yellow, Right=Purple).
So **Option A** is the correct view.
<answer>A</answer>
STEP3-VL-10B
<answer>A</answer>
View original response
<answer>A</answer>
Ground truth
C
View the original prompt
Question
The left image shows the net (unfolded pattern) of a cube, with six faces colored in different colors. The net is folded upward to form a cube. From an axonometric (3D) viewing angle of the cube, which combination of adjacent colors is possible to see? Please answer from options A, B, C, or D.
Options
A. A
B. B
C. C
D. All three other options are incorrect
Ground truth
C
| Model | Answer | Result |
|---|---|---|
| ZDTaichu5.0-9B | C. C | ✓ Correct |
| Qwen3.5-9B | A. A | ✗ Incorrect |
| STEP3-VL-10B | A. A | ✗ Incorrect |
Embodied skill planning
Select the container using the “right” and “nearest” constraints, then plan the pick-and-place steps.
ZDTaichu5.0-9B
The task is to pick up "loki" (the yellow object labeled 3) and place it into the giftbox that is both to the right and nearest. The nearest giftbox to the robot is the green box labeled 1, which is also positioned to the right of the robot's base. Therefore, the required skill sequence is to pick object 3 and place it into container 1, following the pattern ["pick", "place"].
[
{
"name": "pick",
"params": {
"target_entity_name": 3
}
},
{
"name": "place",
"params": {
"target_container_name": 1
}
}
]
View original response
The task is to pick up "loki" (the yellow object labeled 3) and place it into the giftbox that is both to the right and nearest. The nearest giftbox to the robot is the green box labeled 1, which is also positioned to the right of the robot's base. Therefore, the required skill sequence is to pick object 3 and place it into container 1, following the pattern `["pick", "place"]`.
[
{
"name": "pick",
"params": {
"target_entity_name": 3
}
},
{
"name": "place",
"params": {
"target_container_name": 1
}
}
]
Qwen3.5-9B
[
{
"name": "pick",
"params": {
"target_entity_name": 3
}
},
{
"name": "insert",
"params": {
"target_container_name": 2
}
}
]
View original response
```json
[
{
"name": "pick",
"params": {
"target_entity_name": 3
}
},
{
"name": "insert",
"params": {
"target_container_name": 2
}
}
]
```
STEP3-VL-10B
To solve this task, we follow these steps:
- Identify Loki: In the labeled image, the tall yellow figure with a red skirt (Loki) is labeled 3 in the bottom-left quadrant.
- Identify Target Giftbox: The instruction specifies the “right” and “nearest” giftbox relative to the robot arm. In the top-right quadrant (where the robot arm is positioned on the right edge), the green cube labeled 0 is the rightmost and closest to the robot arm, making it the target giftbox.
- Determine Skill Sequence: Placing an object into a container follows the
["pick", "insert"]pattern (valid per the evaluation criteria, as "insert" targets a container).
[
{
"name": "pick",
"params": {
"target_entity_name": 3
}
},
{
"name": "insert",
"params": {
"target_container_name": 0
}
}
]
View original response
To solve this task, we follow these steps:
1. **Identify Loki**: In the labeled image, the tall yellow figure with a red skirt (Loki) is labeled **3** in the bottom-left quadrant.
2. **Identify Target Giftbox**: The instruction specifies the “right” and “nearest” giftbox relative to the robot arm. In the top-right quadrant (where the robot arm is positioned on the right edge), the green cube labeled **0** is the rightmost and closest to the robot arm, making it the target giftbox.
3. **Determine Skill Sequence**: Placing an object into a container follows the `["pick", "insert"]` pattern (valid per the evaluation criteria, as "insert" targets a container).
```json
[
{
"name": "pick",
"params": {
"target_entity_name": 3
}
},
{
"name": "insert",
"params": {
"target_container_name": 0
}
}
]
```
Ground truth
[{"name":"pick","params":{"target_entity_name":3}},{"name":"place","params":{"target_container_name":1}}]
View the original prompt
Question
Put loki into ['right', 'nearest'] giftbox
Ground truth
[{"name":"pick","params":{"target_entity_name":3}},{"name":"place","params":{"target_container_name":1}}]
| Model | Answer | Result |
|---|---|---|
| ZDTaichu5.0-9B | pick(3) → place(1) | ✓ Correct |
| Qwen3.5-9B | pick(3) → insert(2) | ✗ Incorrect |
| STEP3-VL-10B | pick(3) → insert(0) | ✗ Incorrect |
Multi-image and video understanding
Integrate object and scene cues across frames, track events and spatial relationships, and answer questions from a specified position and viewing direction.
Video spatial reasoning in an indoor scene
Integrate indoor video frames to locate the stove from a specified position and viewing direction.
ZDTaichu5.0-9B
The TV is located under the shelf with books (frames 15-16), and the refrigerator is the white appliance in the kitchen (frame 4). The stove is the blue stovetop visible next to the refrigerator (frame 2). When standing by the TV and facing the refrigerator, the stove is in the same general forward direction but positioned to the left side relative to the refrigerator. Thus, it lies in the front-left quadrant.
C
View original response
The TV is located under the shelf with books (frames 15-16), and the refrigerator is the white appliance in the kitchen (frame 4). The stove is the blue stovetop visible next to the refrigerator (frame 2). When standing by the TV and facing the refrigerator, the stove is in the same general forward direction but positioned to the left side relative to the refrigerator. Thus, it lies in the front-left quadrant.
C
Qwen3.5-9B
B
View original response
B
STEP3-VL-10B
B
View original response
B
Ground truth
C
View the original prompt
Question
If I am standing by the tv and facing the refrigerator, is the stove to my front-left, front-right, back-left, or back-right?
The directions refer to the quadrants of a Cartesian plane (if I am standing at the origin and facing along the positive y-axis).
Options
A. back-right
B. front-right
C. front-left
D. back-left
Ground truth
C
| Model | Answer | Result |
|---|---|---|
| ZDTaichu5.0-9B | C. front-left | ✓ Correct |
| Qwen3.5-9B | B. front-right | ✗ Incorrect |
| STEP3-VL-10B | B. front-right | ✗ Incorrect |
Agentic tool use
Break research or computation goals into steps, coordinating search, tool calls, result verification and document generation across multiple turns. Tools are executed by the surrounding application.
These recordings show ZDTaichu5.0-9B working inside the ScienceClaw agent interface, with tool execution and generated artifacts visible alongside the conversation.
Literature review with search and document generation
Search literature on scientific discovery, inspect sources and produce a structured review organized by application area.
ZDTaichu5.0-9B
Demonstrated workflow: The agent searches for relevant papers, inspects source pages, organizes representative work by application area, and produces a structured literature_review.md covering research progress and future directions. The recording shows the tool-call trace and the generated document, illustrating multi-step research, information synthesis, and report creation.
Scientific computation with tools and visual verification
Use mathematical and coding tools to solve a damped oscillator, compare analytical and numerical solutions, and generate plots.
ZDTaichu5.0-9B
Demonstrated workflow: The recording shows tool queries, Python/SciPy computation, checks between the analytical and numerical solutions, and generation of a written report and plots. The final artifacts include the phase-space trajectory and a comparison of displacement and velocity over time, illustrating how the agent connects mathematical reasoning, code execution, result checking, and visualization.
Embodied understanding and action planning
Connect visible scenes, task goals and execution state to identify objects and spatial constraints, plan multi-step actions, and check placement, storage and navigation progress.
These examples combine recorded embodied-task demonstrations with matched-initial-state LIBERO manipulation comparisons. The task setup and observed outcome are described for each example.
Embodied assistant across five tasks
Five task recordings illustrate how scene judgments, target positions and next actions relate throughout a task.
ZDTaichu5.0-9B
This English video plays at 2× speed for approximately 2 minutes 45.7 seconds and brings together five tasks: test tube storage, liquid transfer, bottle rearrangement, utility knife storage and refrigerator navigation. Current judgments, action decisions, image reference points and synchronized left/right gripper-state cues accompany the recorded task sequence.
video chapters and notes
| video chapter (2×, approximate) | Task | Capability focus |
|---|---|---|
| Place both test tubes in the rack | Track tube identity, handoffs and placement | |
| Transfer liquid between cups | Coordinate the dropper and source/destination cups | |
| Move the Sprite bottle to the front | Rearrange objects and reach the target bottle | |
| Store the utility knife in the upper drawer | Open the drawer, transfer the tool and verify storage | |
| Navigate toward the gray refrigerator | Follow visible landmarks and check the route |
Display: Current judgments, action decisions, image reference points and gripper-state cues are synchronized with the recorded task sequence.
These two LIBERO simulation comparisons perform the same two-object storage task from different initial arrangements. Within each group, the three models share the initial state, task, cameras, tools and call budget. These are selected individual runs: ZDTaichu5.0-9B completed the task; STEP3-VL-10B and Qwen3.5-9B did not complete the shown runs.
View each comparison as an inline video at 2× speed, or open the video file or its original-speed MP4.
Soup can and cheese storage · Initial state 2
In initial state 2, store the soup can and cheese box and check placement after releasing and retreating.
- ZDTaichu5.0-9B: Both the cheese box and soup can were placed in the basket and remained stable after release and retreat.
- STEP3-VL-10B: The cheese box was not lifted and the soup can was not placed in the basket before the call budget was exhausted.
- Qwen3.5-9B: The cheese box was stored, but the model then selected the ketchup bottle. The target soup can was not lifted before the call budget was exhausted.
Soup can and cheese storage · Initial state 7
In initial state 7, plan grasps and basket placement from the object arrangement and check the execution outcome.
- ZDTaichu5.0-9B: Both objects were placed in the basket; the gripper released them and retreated upward, completing the storage task.
- STEP3-VL-10B: The cheese box was stored, but the soup can was not securely grasped and remained on the table when the call budget was exhausted.
- Qwen3.5-9B: The cheese box was stored, but the soup can was not lifted. The run ended after four consecutive calls with the same tool and arguments.
Training#
Training recipe
ZDTaichu5.0-9B is trained through a staged BF16 mixed-precision program:
- Vision-language pre-training at 8K: broad perceptual and linguistic alignment.
- Continued pre-training at 16K: richer and more structured multimodal data.
- Supervised fine-tuning at 32K: complex multimodal instructions and step-by-step reasoning.
- Long-context extension to 128K: long documents and long-form video.
- High-quality annealing at 128K: curated instruction and capability-focused data.
- RL from verifiable rewards: a GRPO-based objective using answer correctness, spatial grounding accuracy, and output-format checks, with trivial and unsolvable prompts filtered out.
Data curriculum
The five-stage data curriculum totals approximately 1.28T tokens:
| Stage | Tokens | Focus |
|---|---|---|
| Stage 1 | 364B | Bilingual image-text data, OCR-rich corpora, short-video captions |
| Stage 2 | 737B | Documents, PDF/OCR, long-tail Chinese visual data, captions, video |
| Stage 3 | 136B | Image reasoning, video QA, spatial and physics-aware tasks, CAD-style problems |
| Stage 4 | 28B | Long video and difficult expert instruction data |
| Stage 5 | 10B | Filtered, sampled, and synthetic spatial, OCR, and structured visual tasks |
Acknowledgements#
This model builds on the Qwen3.5 language architecture and NVIDIA C-RADIO vision encoder family. Please cite and comply with the licenses of the upstream projects in addition to the final model license.
Citation#
Use the following project-level citation.
@misc{zdtaichu_5_0_9b,
title = {ZDTaichu5.0-9B: A Multimodal Foundation Model for Visual and Spatial Reasoning, Agents, and Embodied AI},
author = {{ZDTaichu5.0-9B Contributors}},
year = {2026},
note = {Open-weight model and open inference implementation}
}








