StepFun open-sources Step3, a 321B parameter VLM optimized for Chinese AI chips

StepFun has launched Step3, a 321 billion parameter visual language model that activates just 38 billion parameters per token, thanks to custom attention mechanisms like Multi-Matrix Factorization Attention (MFA) and Attention-FFN Disaggregation (AFD). The model, open-sourced on July 31, is positioned as a high-performance alternative to proprietary multimodal models. It scores 74.2 on MMMU and 64.8 on MathVision, marking it as one of the strongest open-access reasoning models available.

Unlike most frontier-scale VLMs, Step3 is designed with inference efficiency in mind. MFA and AFD allow it to cut decoding costs by 4–8 times, making it viable for real-world applications without sacrificing output quality. StepFun's release strategy also focuses on hardware-software co-design. Step3 has been tuned for Chinese AI chips from vendors like Huawei Ascend and Cambricon, a strategic move that aligns with broader efforts to decouple from NVIDIA’s GPU stack in China.

For enterprise developers building multimodal agents, Step3 provides an open and high-performance foundation that can run cost-effectively on local infrastructure. The model's release also signals growing maturity in China’s open-source AI stack, with co-optimization across software and silicon. StepFun is distributing Step3 via Hugging Face, GitHub, and ModelScope under a permissive license.