Skip to main content
Back

TaichuAI's ZDTaichu5.0 Beats Gemini 3 Pro on Spatial Reasoning at 9B

TaichuAI released a 9B open-weight vision-language model that leads spatial reasoning and agent benchmarks against much larger open and closed rivals.

  • TaichuAI released ZDTaichu5.0-9B, an open-weight multimodal model built on Qwen3.5-9B and C-RADIOv4-H vision encoder.
  • Beats Gemini 3 Pro and GPT-5.2 on ViewSpatial, MMSI-Bench, MindCube, and VSI-Bench spatial reasoning benchmarks.
  • Leads 10B-class open models on TAU2-Bench (87.7), Claw-Eval (71.4), and IFEval (93.7) for agentic tasks.
  • Introduces Entropy-Gated Adaptive Recurrent Reasoning that spends extra latent compute on uncertain tokens.
  • Supports 128K context, any-resolution images, video input, and multi-turn tool use out of the box.
  • Ships with a custom vLLM v0.26.0 fork, Docker image, and six community quantizations for Ollama, llama.cpp, and LM Studio.

TaichuAI has released ZDTaichu5.0-9B, an open-weight multimodal model built from a Qwen3.5-9B language decoder and NVIDIA’s C-RADIOv4-H vision encoder. It accepts text, images, and video, supports dynamic visual resolutions, and provides a context window of up to 128K tokens.

The release targets spatial reasoning, embodied understanding, and agent workflows that combine visual input with tool calls. TaichuAI reports unusually strong spatial benchmark results for a model in this size class, including higher scores than several closed models in its comparison tables. The team also provides a customized vLLM branch for deployment.

Spatial scores drive the release

TaichuAI’s benchmark tables show the clearest gains on tasks involving perspective, relative position, mental rotation, and three-dimensional scene understanding.

Benchmark ZDTaichu5.0-9B Named comparison
ViewSpatial 62.50 Gemini 3 Pro: 50.36; GPT-5.2: 47.30
MMSI-Bench 47.20 Gemini 3 Pro: 45.20
MindCube-tiny 78.27 Gemini 3 Pro: 70.87
VSI-Bench 59.69 GPT-5.2: 54.49

All figures come from the project’s release tables. Independent reproduction remains necessary because benchmark versions, prompts, decoding settings, image preprocessing, and video frame selection can materially affect multimodal scores.

The reported agent results include 87.7 on TAU2-Bench, 71.4 on Claw-Eval, and 93.7 on IFEval. For embodied reasoning, the model scores 48 on ERQA and 56 on RoboSpatial, which test scene understanding, affordances, spatial relationships, and interaction-oriented reasoning.

Math and OCR expose the limits

General knowledge and pure mathematical reasoning remain weaker than the strongest closed systems in the release comparison. ZDTaichu5.0-9B scores 77.20 on MMLU-Pro, compared with 87.40 for GPT-5.2 and 89.80 for Gemini 3 Pro. Its AIME 2025 score is 86.70, compared with 100 for GPT-5.2.

The model also trails its Qwen3.5-9B baseline on OCRBench and MMStar. Document extraction, dense OCR, and competition mathematics therefore require workload-specific testing before adoption.