Step3-VL-10B
Open 10B vision-language model that StepFun says rivals models 10 to 20 times larger, such as GLM-4.6V and Qwen3-VL-Thinking.
Unified pretraining plus parallel coordinated reasoning at test time (PaCoRe). StepFun publicly corrected errors in the baseline numbers it reported for Qwen3-VL-8B. Apache-2.0.
- Date
- Tuesday, 13 January 2026
- Lab
- StepFun
- Kind
- open-weights
- Access
- open weights
HF weights uploaded 2026-01-13 UTC; the model card links arXiv 2601.09668. The card says baseline metrics (AIME, HMMT, LCB for Qwen3-VL-8B) were wrong due to a max-token setting and are being re-run.
Sources
This record was checked against its sources on 6 October 2026. How we check