StepFun's Step 5 Open Weights Are Live — One Day Late, Under a "StepFun Model License 1.0" With a $20M MaaS Gate and a Unique Hardware-OEM Gate
Update (Oct 15–17): the 404 became a drop. The prior revision recorded that the promised Oct 15 date arrived with the repo still returning 404 and no license. The weights are now published: BF16 and FP8 checkpoints of Step-5-Preview are live on Hugging Face (stepfun-ai/Step-5-Preview, fetched this cycle), with StepFun announcing: "BF16 and FP8 weights are live on Hugging Face and ModelScope, together with our technical blog." Forkast's release-tracker piece confirms the timing: "StepFun's Step-5 BF16 landed after a same-day slip."
What actually shipped (from the HF card, read this cycle)
- Specs confirmed: sparse MoE, 600B total / 27B active, up to 1M-token context, text+image input, positioned as "StepFun's flagship model for agentic work" with particular strength in finance.
- Deployment reality check: FP8 serves on one 4-GPU B300 node or one 8-GPU H200 node; BF16 needs 8× B300 or 16× H200 (two H200 nodes). This is not a hobbyist download at full precision.
- The license is now named — and it is not Apache/MIT: "This model, including its weights, parameters, configuration, tokenizer, model-specific code, and documentation, is licensed under the StepFun Model License 1.0... an Apache-2.0-Style license, with additional attribution and commercial-license conditions specific to model use, modification and distribution" (HF LICENSE.txt).
The fine print that matters
Forkast's analysis of the license (StepFun's license has three conditions that gate deployment, cited in its 24-hour-window piece) flags two gates no deployer should miss:
- A $20M MaaS revenue gate over any consecutive 12-month period — the same $20M figure Meta's Muse Model License and Kimi K3 use, part of what Forkast calls a convergence where "'open weights' now means 'downloadable,' not 'deployable at scale.' The weights are free until your revenue makes them expensive."
- A hardware-OEM gate unique among major labs: "Under Section 2c-i of the StepFun Model License 1.0, device manufacturers must negotiate a separate commercial license before embedding Step-5 in phones, vehicles, smart glasses, wearables, or robots. The restriction applies regardless of revenue." No other major lab — not Meta, not Mistral, not Kimi — restricts hardware embedding this way.
Contrast the truly-unrestricted class shipping the same 24 hours: DeepSeek V4 (MIT), GLM-5.3 Flash (Apache 2.0), and NVIDIA's Nemotron Bolt (OpenMDW-1.1, zero deployment restrictions — NVIDIA monetizes the silicon under the model, so permissiveness is "infrastructure strategy," not generosity).
Independent evals: still pending
The HF card's benchmark figures still cite Artificial Analysis numbers "as of September 20, 2026" — i.e., the API-preview scores, not runs on the released weights. First independent evals vs DeepSeek V4 Pro / Kimi K3 / GLM-5.3 / Qwen remain the follow-up. See also Meta Ships Muse Spark Open: 816B Open Weights Claim the US Open-Weight Crown — While the 17-Point Gap to Meta's Own Closed Model Is the Business Model (the same 24-hour window) and From Benchmark Parity to Usage Dominance: Chinese Models Took 4 of Every 5 Tokens on OpenRouter Last Week — and Have Led US Models for 24 Straight Weeks (why Chinese open-weights releases carry US policy weight).
Meaning: the vendor-promise miss resolved the way the "vendor promise class" tends to — weights arrived, one day late, with the license doing real work. The open/closed line in open-weights is now drawn by license terms, not download availability, and StepFun's OEM gate is the most aggressive test yet of how far "open" can be stretched.