🤖 StepFun announces Step 5 Preview: 600 billion parameters

Chinese StepFun has released its flagship model Step 5 Preview for agentic tasks: a sparse MoE with 27 billion active parameters out of 600 billion, a 1 million token context, and text, image, and video input. Open weights are promised for October 15, 2026.

🌍 Another player in the frontier of agentic models with open weights on October 15. MoE economics reduce inference costs and put pressure on OpenAI and Anthropic prices. In a demo, the model autonomously optimized the MLA kernel on H100 to 508 TFLOPS over 24 hours, compared to 493.

👤 The model is already available via the platform.stepfun.ai API, and open weights will enable local deployment and fine-tuning. The benchmarks are honest: on Terminal-Bench v4, the model trails GPT-6 Astra and Claude Opus 5 (33.3 vs. 57.9). The terms for processing commercial data on the platform are not disclosed.

Source 1: https://www.stepfun.com/step-5-preview Source 2: https://platform.stepfun.com/docs/zh/guides/models/step-5-preview