ZDTaichu 5.0-9B Open-Sourced: 9B Multimodal Model Tops 8 of 9 Spatial Benchmarks

机器之心 · wechat · 2026-09-17

Zidong Taichu has open-sourced ZDTaichu 5.0-9B, a general multimodal model focused on spatial embodied intelligence. It ranks first in its parameter class on 8 of 9 international spatial understanding benchmarks, beating Qwen3.5-9B and STEP3-VL-10B, and still scores highest on 5 benchmarks even against closed models like Gemini 3 Pro and Grok 4, while keeping first-tier general multimodal abilities (OCR, code, agents).

Adaptive recurrent reasoning: after a standard forward pass, an entropy gate assesses uncertainty — confident predictions output directly, while hard spatial judgments iterate layer-block computation internally in latent space, unlike CoT-style longer outputs. Stability comes from damped updates, dual stop criteria (KL divergence + hidden-state residual), and trajectory readout with state rollback. The team notes this resembles the rumored Loop Transformer approach of GPT-6 Astra.
Open data pipeline: beyond weights, the team published detailed data production recipes spanning pretraining, SFT, annealing, and GRPO RL — including dedup/filtering, instruction verification, a 3D capability-difficulty-quality labeling system, and verifiable rewards like spatial box-hit scoring — so companies can continue training on their own industrial and robot data.
Real-world validation: closed-loop tested in research automation (question-to-report workflows, wet-lab tube tracking) and smart manufacturing (cross-station material delivery, machine tending). The 10B size targets edge deployment constraints, and the open release signals spatial embodied AI competition shifting from benchmarks to data loops and ecosystem building.

Original post →

ZDTaichu 5.0-9B Open-Sourced: 9B Multimodal Model Tops 8 of 9 Spatial Benchmarks

More from Models