StepFun Step 5 Preview open weights: 600B MoE, read the license
StepFun posted Step 5 Preview open weights: a 600B MoE with 27B active and a 1M-token context. It needs 4 B300s or 8 H200s, and the license has conditions.
StepFun put the Step 5 Preview open weights on Hugging Face this week, and the spec sheet reads like a frontier agent model. It is a sparse mixture-of-experts with 600B total parameters, 27B active per token and a 1M-token context window. The model card calls it StepFun's "flagship model for agentic work." Two facts matter more than the benchmarks: the hardware it needs, and the license it ships under.
What actually happened
StepFun announced that BF16 and FP8 weights are live on Hugging Face and ModelScope, with a technical blog. The API version came first. The weights followed on October 15. The details from the model card:
- Size. 600B total, 27B active. Text and image input, text output. Up to 1M tokens of context, input and output combined.
- Built for long tool loops. The card targets tasks that need repeated tool calls, execution, verification and revision.
- Hardware. For FP8 on vLLM or SGLang: one 4-GPU B300 node or one 8-GPU H200 node. BF16 doubles that. This is not a Mac Studio model.
- Serving. vLLM and SGLang, both with OpenAI-compatible APIs. NVIDIA says it also published an NVFP4 checkpoint.
- Scores. The card reports 67.7 on DeepSWE and 33.3 on Terminal-Bench. StepFun ran these numbers itself.
- License. StepFun Model License 1.0. The card calls it "Apache-2.0-style" with "additional attribution and commercial-license conditions." That is not Apache 2.0.
Why open-weight models matter for your business
Open weights give you leverage, even if you never run them. Each serious open release with an OpenAI-compatible API puts a price ceiling on the hosted models you pay for today. That is the real win for a small team.
But "open weights" is not "free to use any way you like." A license that adds commercial conditions to Apache-style terms can hide user caps, attribution rules or a separate license for some uses. The card does not list those terms. They are in the license file. Read the file before a client's product depends on this model, not after.
Then do the hardware math. Eight H200s is a cluster, not a side project. For most operators the sensible path is a hosted provider that serves the open weights, with your evals on your tasks, and the weights as a fallback you could move if the price or terms change.
Key takeaways
- Step 5 Preview is a 600B MoE with 27B active parameters and a 1M-token context
- FP8 serving needs one 4-GPU B300 node or one 8-GPU H200 node
- It ships under the StepFun Model License 1.0, not Apache 2.0, with extra commercial conditions
- Benchmark scores are StepFun's own. Test it on your tasks before you trust them
- Its real value for most teams is price pressure and a portable fallback
Open weights are an exit, if your stack can use one. We build model layers that run on hosted APIs today and move to open weights when the numbers say so, with your evals as the gate. See how we build it, or tell us what you run.
Sources: StepFun on Hugging Face, StepFun on X, NVIDIA AI on X.
- #open-weight-models
- #stepfun
- #self-hosting
- #ai-agents
- #model-licensing
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Opsin raises $21M for AI agent security: 93% had full access
Opsin raised $21M for AI agent security after finding 93% of misaligned agent tools had full system access and 1 in 200 needed approval. Scope yours now.
Read itAI insurance exclusions: Marsh CEO warns, read your renewal
Marsh's CEO says broad AI insurance exclusions create confusion. Before your agents act on your behalf, check what your liability policy says about AI.
Read it