← October 15, 2026 briefing · added in the 20:00 KST update
StepFun releases Step 5 Preview open weights: 600B total / 27B active MoE, with tech blog on 92-layer, Sparse GQA design (report)
Announced October 15, 2026
Verified (fact)
StepFun's Step 5 Preview, which first appeared on OpenRouter on the 10th, has been released as open weights. It is a MoE model with 600B total parameters and 27B active parameters, and a context window of 1 million tokens. An accompanying technical blog post describes its 92-layer architecture, Sparse GQA, and its training approach for long-running agents.
Needs fact-check (claim)
StepFun claims that, as a flagship model for agentic tasks, it delivers frontier-level performance in software engineering and professional knowledge work, and is especially strong in finance (as quoted in a Reddit post).
Why it matters (interpretation)
With the weights of a previously API-only model now public, developers can serve or fine-tune a 1M-token-context MoE model themselves. At 600B parameters, self-hosting requires multi-GPU infrastructure.
Confidence medium · official source pending
Sources
Story thread
- 2026-10-10 StepFun releases 'Step 5 Preview,' a 600B MoE with a 1M-token context, on OpenRouter (community report)
- 2026-10-15 StepFun releases Step 5 Preview open weights: 600B total / 27B active MoE, with tech blog on 92-layer, Sparse GQA design (report)
More about StepFun
-
StepFun unveils its first 'agent phone' that shows AI before apps on unlock, priced at 3,999 yuan (report)
According to InfoQ China, Chinese AI company StepFun (阶跃星辰) released its first agent smartphone, priced at 3,999 yuan. Unlocking the...
-
StepFun releases 'Step 5 Preview,' a 600B MoE with a 1M-token context, on OpenRouter (community report)
According to GeekNews, StepFun's 'Step 5 Preview' has been listed on OpenRouter. It is a flagship model for agentic work such as...