StepFun's Step 5 Preview Lands on OpenRouter With Open Weights Due Oct 15
StepFun's Step 5 Preview, the company's flagship for agentic work, is now listed on OpenRouter: a 600-billion-parameter sparse MoE with 27 billion active parameters, a 1M-token context window, and text, image, and video input. Open weights are scheduled for release on 15 October, which would make it one of the largest multimodal open models available for self-hosting.
On this page
The listing, and what the model is
StepFun's Step 5 Preview, the company's flagship model for agentic work, is now available through OpenRouter: a sparse Mixture-of-Experts with 600 billion total parameters and 27 billion active per token, a 1M-token context window with up to 64K output tokens, and text, image, and video input. The listing also appears on Vercel's AI Gateway, with pricing reported around a dollar per million input tokens.
The date that matters for this audience is 15 October: StepFun has scheduled the open-weights release for then, which would make Step 5 one of the largest multimodal open models ever self-hostable, alongside Reflection's Beam and Xiaomi's MiMo line. The model itself launched in September; what is new this week is Western API availability and the weights commitment moving onto a calendar.
Where it sits in an increasingly crowded tier
Step 5 enters a tier that did not exist a quarter ago: frontier-adjacent, agentic-first models with 1M contexts and announced open weights. Reflection's Beam claims 77.2 on SWE-Bench Pro v2-Hard at 23B active; Mistral Large 4 previews at 49B active with open weights due this month; Xiaomi's MiMo-V2.6 line ships 1M-context checkpoints with its RL methodology published. Third-party coverage of Step 5 reports strong coding scores, roughly 67 of 80 on the coding benchmarks its listing cites, and early comparisons position it as a serious open-weights contender rather than a flagship-priced novelty.
For buyers, the practical difference between these models is narrowing to what surrounds the weights: license terms, inference efficiency, tool-calling quality, and whether the training and evaluation methodology is published. StepFun has not yet named the license for the 15 October release, which is the same unanswered question that hung over Mistral's announcement.
Watch the weights, not the preview
The pattern across this month's releases is consistent: API preview first, weights on a dated promise later. Beam, Mistral Large 4, and now Step 5 have all made open-weights commitments with the model still behind a gate. Two of the three have named dates; the third shipped early in the month. The preview-to-weights gap is where announcements historically go quiet, so the 15 October date is the thing to check, not the benchmark table.
When the weights land, the local-hosting question will be the usual one: a 600B sparse model with 27B active is an expert-caching or multi-GPU proposition, in the same class as Beam and Mistral Large 4 rather than a laptop model. But the tier of self-hostable, agentic, long-context, multimodal models now has three entrants from three countries, and every one of them is priced or licensed for people who run their own infrastructure. That competition is the story, and October is the month it got real.