ZDTaichu5.0-9B: An Open 9B Multimodal Model Built for Spatial Reasoning

ZDTaichu5.0-9B is an open multimodal foundation model from TaichuAI aimed at visual understanding, spatial reasoning, agentic tool use, and embodied-AI research. Its practical appeal is not simply that it can interpret images: the project is explicitly designed for tasks where a system must reason about viewpoint changes, object relationships, multi-image evidence, and video over long contexts. 2

Architecture: language decoder plus vision encoder

The model combines a roughly 9-billion-parameter Qwen3.5-9B language decoder with a C-RADIOv4-H vision encoder. It accepts text, single or multiple images, and video; the release describes support for any-resolution visual inputs and a context window of up to 128K tokens. 2

That input range makes it relevant to more than ordinary image captioning. The model card positions ZDTaichu5.0-9B for documents, charts, diagrams, OCR, visual question answering, and visual mathematics, alongside broader scene interpretation. 2

What it is designed to understand

TaichuAI emphasizes spatial and embodied tasks that are often difficult for general-purpose vision-language models. The stated capability set includes:

• 2D and 3D spatial reasoning, including object relations and scene structure
• Viewpoint changes and mental transformation, where the reference frame changes
• Multi-view association and cross-view comparison
• Video and multi-image understanding, including event tracking and retrieval of details from longer footage
• Embodied-AI concepts, such as spatial perception, affordances, manipulation understanding, and high-level action planning 2

The model also supports agent-oriented interactions. But this should not be confused with autonomous execution: the model can plan tool calls and participate in multi-step or multi-turn workflows, while the host application remains responsible for executing, validating, and securing any tools. 2

Entropy-gated recurrent reasoning, explained

ZDTaichu5.0-9B’s central technical claim is Entropy-Gated Adaptive Recurrent Reasoning. Instead of applying the same additional reasoning computation to every token, the system uses uncertainty to decide where extra latent-space refinement is warranted. 2

In plain terms, straightforward steps can proceed normally, while uncertain or difficult predictions can receive additional internal refinement. The intended trade-off is greater effective computational depth on hard reasoning steps without uniformly increasing work across the entire response. 2

This is particularly relevant to spatial questions, where a single ambiguous relation—such as a changed camera angle or an object’s position relative to another—can determine whether the final answer is correct.

Reported benchmark results

TaichuAI reports the following results in its model materials:

Evaluation area	Benchmark	Reported score
Spatial / embodied	ViewSpatial	62.50
Spatial / embodied	MMSI-Bench	47.20
Spatial / embodied	MindCube-tiny	78.27
Spatial / embodied	ERQA	48.00
Spatial / embodied	RoboSpatial	56.00
Agent / instruction following	TAU2-Bench	87.70
Agent / instruction following	Claw-Eval	71.40
Agent / instruction following	IFEval	93.70

2

The project positions the model as a leader in spatial and embodied reasoning among the roughly 10B-scale general-purpose VLMs included in its comparisons, while retaining strong general visual capabilities. 2

That is a useful signal for practitioners selecting a compact open model for spatially demanding workloads. It is not, however, a universal ranking. The reported comparisons depend on the vendor’s chosen model versions, prompts, preprocessing, decoding settings, and evaluation setup. Independent reproduction with those details disclosed is needed before treating the results as settled comparative performance. 2

Availability, licensing, and deployment

ZDTaichu5.0-9B is available through TaichuAI’s Hugging Face repository. The release identifies the model as custom-code based and links project materials for recurrent reasoning and deployment. 2

The stated licensing arrangement is the NVIDIA Open Model License for the released model materials, with Apache-2.0 notices for Qwen3.5 components. Teams considering redistribution or commercial use should review the repository’s current license files and notices directly. 2

For serving, TaichuAI documents custom vLLM 0.26.0 and Docker paths, including different sampling presets for spatial-grounding workloads and general tasks. 2

The bottom line

ZDTaichu5.0-9B is an open, approximately 9B-scale multimodal model built around a clear specialization: reasoning across images, viewpoints, scenes, and video rather than only recognizing visual content. Its 128K context, any-resolution visual input, and adaptive recurrent-reasoning approach make it a notable option for researchers and developers exploring spatial VLMs, embodied AI, and host-managed agent workflows. 2

Its published results are promising, especially for spatial benchmarks, but they should be evaluated as vendor-reported evidence—not as independently established performance—until third parties reproduce the tests under transparent conditions. 2