ZDTaichu5.0-9B Ships Under an NVIDIA Licence Its Hugging Face Tags Omit, and Its Agent Scores Mix Test Setups
A.I. / news
ZDTaichu5.0-9B Ships Under an NVIDIA Licence Its Hugging Face Tags Omit, and Its Agent Scores Mix Test Setups
TaichuAI's 9-billion-parameter vision-language model leads its open-weight peers on spatial benchmarks in the developer's own tables. The licence and the judge behind the agent scores matter to anyone deploying it.
TaichuAI has released ZDTaichu5.0-9B, a 9-billion-parameter model for images, video, spatial reasoning and tool use, with weights downloadable from Hugging Face without a gate. The repository showed 12,139 downloads and 2,080 likes on Wednesday, and its last modification is dated September 20.
The licence is the first thing to check. The model card says the weights are "made available under the NVIDIA Open Model License Agreement, with the Qwen3.5 Apache-2.0 license and all other third-party notices retained." The repository's metadata tags carry no licence entry, so a search filtered by licence will not surface it.
Why an NVIDIA licence on a Qwen model
The card says the model pairs a Qwen3.5-9B language decoder with NVIDIA's C-RADIOv4-H vision encoder. That encoder is the reason NVIDIA's terms apply. The licence text in the repository says NVIDIA models under it are "commercially useable" and that users are "free to create and distribute Derivative Models."
The same text reserves a right that Apache 2.0 does not. It says NVIDIA may update the agreement to comply with legal and regulatory requirements at any time, and that a user must either follow the updated licence or stop using the model and any derivative. It also ends the licence if a user files a copyright or patent suit over the model. The card asks users to comply with upstream licences "in addition to the final model license."
Licence mismatches on Qwen-based releases are a recurring problem: we reported on Edge0's Audio8 and its Qwen licence gap, and Qwen's own Qwen3.8-27B ships under Apache 2.0.
The context window is up to 128K tokens. The GitHub instructions call for CUDA 12.9 or later and driver 575.51.03 or later, and list FP8 and NVFP4 variants alongside the standard weights.
What the benchmark tables show
The card claims the model "leads spatial capability among the compared 10B-scale general-purpose VLMs." All figures below are from TaichuAI's own tables. No independent party is named as having run them.
| Benchmark | ZDTaichu5.0-9B | Qwen3.5-9B | Gemini 3 Pro |
|---|---|---|---|
| ViewSpatial | 62.50 | 48.20 | 50.36 |
| MindCube-tiny | 78.27 | 57.60 | 70.87 |
| ERQA | 48.00 | 41.50 | 66.00 |
| TAU2-Bench | 87.70 | 79.10 | 85.40 |
The gain over its own language backbone is largest on ViewSpatial, 14.3 points. On ERQA, which covers embodied reasoning, it trails Gemini 3 Pro by 18 points, so the claim of leadership holds only inside the 10B class.
- ZDTaichu5.0-9B87.7 points
- GPT-5.287.1 points
- Gemini 3 Pro85.4 points
- STEP3-VL-10B81.7 points
- Qwen3.5-9B79.1 points
Source: ZDTaichu5.0-9B model card, accessed 2026-09-30
The agent scores come with a footnote
The agent claims are TAU2-Bench at 87.7 and Claw-Eval at 71.4. The card's footnote says the local runs "use DeepSeek-V4-Flash-0731 as the simulated user and/or judge," while externally reported scores "follow the evaluation setup of their cited sources."
That means the TAU2-Bench row sets a locally judged number beside scores produced under other setups. The 0.6-point gap to GPT-5.2's 87.1 is inside what a different judge could move.
The multi-image spatial tests also used a modified prompt. The card says the evaluation prompt added a requirement that reasoning sit inside tags and the final answer in a boxed format. The card also says the model does not run tools itself, and that the surrounding application must implement and secure them.
The card names no company behind TaichuAI, and its citation credits only "ZDTaichu5.0-9B Contributors." The card does not say when a technical report will appear.
Sources
More in A.I.
- 01GPT-Synopsys: OpenAI Gets Paid Only When the Chips It Helps Design Beat the Customer's BaselineThe Sept. 30 deal pairs an OpenAI model with Synopsys' design software, with no price, no release date and no named customer.
- 02Ataraxos Beats Stratego's Top Player 15-1-4 After Training on 16 H100s for a WeekA Nature paper from MIT, Carnegie Mellon, NYU and Stanford puts the compute bill under $8,000, against an estimated $3 million to $4.5 million for DeepMind's DeepNash.
- 03Nine Mathematicians Advising OpenAI Ask AI Labs to Stop Testing Hard Problems on Models Nobody Else Can UseThe Advisory Group on Mathematics and AI published its rules on Sept. 29: release fast, fund human understanding, disclose prompts and costs.
- 04Qwen3.8-27B Ships Under Apache 2.0 and Fits in 17GB, but Spends 160 Million Tokens Where the Median Spends 43 MillionAlibaba's open-weight model scores 52 on Artificial Analysis's Intelligence Index. Its own benchmark figures are vendor-supplied, and users report slow runs.