MODEL RELEASE

ZDTaichu5.0-9B

From Visual Understanding to Spatial Intelligence

MULTIMODAL FOUNDATION MODEL

ZDTaichu5.0-9BVisual · Spatial · Embodied · Agent
9B128K CONTEXTTEXT · IMAGE · VIDEO

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.

  1. 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.
  2. Refine states in latent space: Damped, anchored updates progressively refine hidden states, allocating additional computation to difficult tokens.
  3. Stop using convergence signals: KL divergence and hidden-state residuals guide stopping, dynamically controlling the depth of recurrent computation.
  4. Select an output from the trajectory: Trajectory readout selects the lowest-risk state, with rollback when needed.

Benchmark Results#

Comparison with open models

ZDTaichu5.0-9B benchmark comparison with open modelsSVG ↓

Comparison with closed models

ZDTaichu5.0-9B benchmark comparison with closed modelsSVG ↓

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.

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.

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.

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.

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.

Training#

Training recipe

ZDTaichu5.0-9B is trained through a staged BF16 mixed-precision program:

  1. Vision-language pre-training at 8K: broad perceptual and linguistic alignment.
  2. Continued pre-training at 16K: richer and more structured multimodal data.
  3. Supervised fine-tuning at 32K: complex multimodal instructions and step-by-step reasoning.
  4. Long-context extension to 128K: long documents and long-form video.
  5. High-quality annealing at 128K: curated instruction and capability-focused data.
  6. 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}
}
Download ↗