ZDTaichu 5.0-9B open-sourced, topping 8 of 9 spatial understanding benchmarks
量子位 · wechat · 2026-09-16
Taichu AI open-sourced ZDTaichu 5.0-9B, a 9B multimodal model focused on spatial understanding and embodied tasks. It ranks first on 8 of 9 international spatial benchmarks, scoring 78.27 on MindCube-tiny vs Gemini3Pro's 70.87 and Grok4's 63.56, while keeping general skills strong (AI2D 91.48, MathVistaMini 84.5, OCRBench 85.5).
- Demoed on target localization, cross-viewframe reasoning, and robotic task planning (box收纳, test-tube transfer, part loading).
- Training uses a three-stage data pipeline (pretraining alignment, SFT with embodied/tool-calling traces, annealing + GRPO with verifiable rewards for coordinate hits and format compliance).
- Inference adds adaptive recurrent computation: entropy gating identifies uncertain tokens and iterates partial layers internally, stabilized by damped updates, dual stop criteria, and state rollback—mirroring OpenAI GPT-6 Astra's adjustable reasoning spend.
- Weights and the full data-engineering pipeline are open-sourced, letting enterprises retrain with their own footage and simulation data.
- Links: Blog · GitHub · HuggingFace
Related event: Zidong Taichu Open-Sources 9B Multimodal Model, Tops Spatial Benchmarks(3 posts)→
More from Embodied
- ANYbotics begins customer deliveries of ANYmal X, first legged robot certified for explosive atmospheres — lukas_m_ziegler · 2026-11-02
- RAM prices tripled, so Tesla cut a third of memory from Optimus's brain — aakashgupta · 2026-11-02
- Too lazy on weekends? Watch this robot tidy up a desktop — flngr · 2026-11-02
- Wearing Whoop and Apple Watch on a single dual-slot band — tinyfool · 2026-11-02
- REA: async brain-cerebellum runtime for robots where policies never wait on the LLM — Pengfei Zhou · 2026-11-02
- Robot Combat Sports Coming to Japan with a Permanent Arena in Tokyo — cixliv · 2026-11-02