Step3-VL-10B

Step3-VL-10B is StepFun's lightweight open-source foundation model for multimodal intelligence, released in early 2026. Despite its compact 10B footprint, StepFun claims it matches or beats significantly larger open-weight models on visual perception, complex reasoning, and human alignment, including 94.43% on AIME2025 and 80.11% on MMMU per third-party reporting.

It uses Parallel Coordinated Reasoning (PaCoRe) to scale test-time compute and was trained on 1.2T multimodal tokens with over a thousand reinforcement-learning iterations. License details on the Hugging Face card should be checked before commercial use.