What is ZDTaichu5.0-9B?
ZDTaichu5.0-9B is an open vision-language model from Zidong Taichu, the model line built by the Institute of Automation of the Chinese Academy of Sciences (CASIA) with the Wuhan Institute of Artificial Intelligence. It was open-sourced on 15 September 2026. The model card does not name the institutes; the attribution comes from Chinese press and the Wuhan science bureau, and the official Docker image sits on the Wuhan institute's registry. It pairs Qwen3.5-9B as the language model with NVIDIA's C-RADIOv4-H image encoder, 9.79 billion parameters in all, and reads text, single or multiple images and video into a 128K-token context.
The pitch is spatial and embodied reasoning: left and right, viewpoints, 3D layouts, what a robot can grasp, plus multi-step tool use. Its one architectural addition, Entropy-Gated Adaptive Recurrent Reasoning, reruns a block of layers on tokens where the model is uncertain and stops when the hidden state settles, so hard tokens get more compute without a longer written chain of thought. Training ran to about 1.28 trillion tokens across five stages, finished with GRPO reinforcement learning.
Weights come in BF16, FP8, NVFP4 and GGUF, with a 3.2B DSpark draft model for speculative decoding. The code on GitHub is Apache 2.0, but the weights are under the NVIDIA Open Model License because of the C-RADIO encoder. Adoption is small so far: 2,108 Hugging Face downloads of the main repo in the last 30 days and 2,382 for the official GGUF, against 3,043 likes and 3,063 GitHub stars.
Is ZDTaichu5.0-9B worth it?
ZDTaichu5.0-9B is worth a look only if your images are about space rather than text. In the team's own table it scores 62.50 on ViewSpatial against 48.20 for its own base, Qwen3.5-9B, 78.27 on MindCube-tiny against 57.60 and 47.20 on MMSI-Bench against 38.70, ahead of Gemini 3 Pro on all three. Every figure here is self-reported. We found no entry on Artificial Analysis or OpenCompass, and the multi-image spatial tests were run with an added prompt asking for step-by-step reasoning in think tags.
The cost of that specialisation shows on ordinary vision work. Against the same Qwen3.5-9B it loses on OCRBench (85.50 to 89.20), MMStar (76.80 to 79.70), RealWorldQA (76.99 to 80.30) and MathVerse (76.40 to 84.14), and on MMLU-Pro knowledge it drops from 82.50 to 77.20. For invoices and screenshots, the base model it was built from is the better pick by its makers' own count.
The embodied and agent numbers are better but narrower than the headline. ERQA at 48.00 and RoboSpatial at 56.00 lead the 10B-class open models listed, yet sit well under Gemini 3 Pro's 66.00 on ERQA. TAU2-Bench at 87.70 and Claw-Eval at 71.40 are the release's boldest claims, and both were scored locally with DeepSeek-V4-Flash as the simulated user or judge, so they are not directly comparable with figures published elsewhere. IFEval at 93.70 and LiveCodeBench v6 at 73.40 against Qwen's 65.60 suggest the post-training did help text tasks.
Running it is less simple than the 9B size implies. The weights load only with trust_remote_code, and serving needs Taichu's own vLLM 0.26.0 branch or a Docker image hosted on a Wuhan institute registry, with CUDA 12.9 or later. The licence is the second thing to read: commercial use is allowed, but it is NVIDIA's, it is revocable if you sue over patents, NVIDIA may change it, and every redistribution must carry a set notice. A GitHub account created on 4 September 2026 now shows 3,063 stars, more than the main repo's 2,108 monthly downloads, which says more about launch attention than about use in production.
Where it earns its place
A robotics researcher at a university lab building a vision-language-action policy
Start from a small model that already understands where things are in a scene
Fine-tune ZDTaichu5.0-9B rather than a general 9B, since its self-reported ViewSpatial, VSI-Bench (59.69) and RoboSpatial scores beat Qwen3.5-9B. At 9.79B parameters it fits one 24 GB GPU for inference; check the NVIDIA licence notice before publishing a derived model.
An ML engineer at a warehouse-automation firm
Answer layout questions from fixed cameras, such as which shelf is behind which, on premises
Serve the FP8 build (12.67 GB of weights) through Taichu's vLLM image on one GPU, with temperature 0 as the card recommends for spatial tasks. Validate on your own footage first: the spatial lead is the vendor's own measurement.
A platform team lead at a 200-person software company
Run a cheap local agent model that calls internal tools and also reads screenshots
Enable the qwen3_coder tool-call parser in vLLM and test against your own tool suite, because the 87.70 TAU2-Bench score was judged by DeepSeek-V4-Flash. The licence allows commercial use with no revenue cap, and the hardware is a single GPU rather than per-token fees.
Before you commit
- Setup
- pip install transformers 5.3.0, torch 2.10.0 and timm, then load with AutoModel and trust_remote_code. For an OpenAI-compatible server, use the Taichu vLLM v0.26.0 Docker image or branch, which needs CUDA 12.9 or later and NVIDIA driver 575.51.03 or later.
- Integrations
- Hugging Face and ModelScope downloads; vLLM (forked), Transformers, and llama.cpp or Ollama through the GGUF files with an mmproj vision file. Reasoning and tool calls use vLLM's qwen3 and qwen3_coder parsers; the DSpark draft model plugs into vLLM speculative decoding.
- Data and privacy
- Self-hosted, so images stay on your hardware. Weights are under the NVIDIA Open Model License: commercial use and derivatives allowed, no revenue cap, but the licence ends if you bring patent or copyright claims over the model, NVIDIA may update it, and redistribution must include the licence and its notice. Qwen3.5 parts and the repo code are Apache 2.0.
- Cost at scale
- Weights are free and no hosted price exists. Weight files are 19.59 GB at BF16, 12.67 GB at FP8 and 9.81 GB at NVFP4; the GGUF Q4_K_M is 5.63 GB plus a 0.88 GB Q8_0 vision file. NVFP4 is native only to NVIDIA Blackwell GPUs. The 6.42 GB DSpark draft model gave 2.46x to 3.17x speed-ups on one H800 in Taichu's tests.
- Limits
- 128K-token context on the card. Only 8 of 32 language layers use full attention, the rest linear attention, which keeps long-context memory down but is why stock vLLM does not run it. Videos are sampled to a set frame count (8 in the card's example). No independent evaluations yet.
What’s good
- 62.50 on ViewSpatial and 78.27 on MindCube-tiny, far ahead of Qwen3.5-9B (self-reported)
- Leads listed 10B-class open models on ERQA and RoboSpatial
- BF16, FP8, NVFP4 and GGUF builds plus a DSpark draft model from launch
- Commercial use allowed; fits a single 24 GB GPU at BF16
What’s not
- Trails its own Qwen3.5-9B base on OCRBench, MMStar and MMLU-Pro
- No independent benchmark results yet; agent scores judged by DeepSeek-V4-Flash
- Needs remote code and a forked vLLM build
- NVIDIA Open Model License, not Apache 2.0, for the weights
Best for
Robotics and embodied-AI researchers who need a small open model that reasons about viewpoints, layouts and affordances, and can live with a forked serving stack.
ZDTaichu5.0-9B alternatives
ZDTaichu5.0-9B: common questions
- Is ZDTaichu5.0-9B worth it?
- A specialist dressed as a generalist: on its makers' own numbers it reads rooms and robot scenes far better than Qwen3.5-9B, and reads text slightly worse. We rate it 3.8 out of 5.
- How much does ZDTaichu5.0-9B cost?
- Free weights under the NVIDIA Open Model License, with the Qwen3.5 parts under Apache 2.0. No hosted API or per-token price is published; you pay for your own GPU.
- What are the downsides of ZDTaichu5.0-9B?
- Trails its own Qwen3.5-9B base on OCRBench, MMStar and MMLU-Pro; no independent benchmark results yet; agent scores judged by DeepSeek-V4-Flash; needs remote code and a forked vLLM build; nVIDIA Open Model License, not Apache 2.0, for the weights.
- Who is ZDTaichu5.0-9B best for?
- Robotics and embodied-AI researchers who need a small open model that reasons about viewpoints, layouts and affordances, and can live with a forked serving stack.
- How much does ZDTaichu5.0-9B cost for a team or in production?
- Weights are free and no hosted price exists. Weight files are 19.59 GB at BF16, 12.67 GB at FP8 and 9.81 GB at NVFP4; the GGUF Q4_K_M is 5.63 GB plus a 0.88 GB Q8_0 vision file. NVFP4 is native only to NVIDIA Blackwell GPUs. The 6.42 GB DSpark draft model gave 2.46x to 3.17x speed-ups on one H800 in Taichu's tests.
- Is ZDTaichu5.0-9B safe for company data?
- Self-hosted, so images stay on your hardware. Weights are under the NVIDIA Open Model License: commercial use and derivatives allowed, no revenue cap, but the licence ends if you bring patent or copyright claims over the model, NVIDIA may update it, and redistribution must include the licence and its notice. Qwen3.5 parts and the repo code are Apache 2.0.
- What are the best ZDTaichu5.0-9B alternatives?
- The closest alternatives we have reviewed are MiniCPM-V 4.7 (35B-A3B and 1B), Gemma 4 and Phi-4-multimodal.
Answers checked Oct 31, 2026.