Step 5 Preview's open weights arrive — with a bespoke license
StepFun published the weights for Step 5 Preview on Thursday, releasing BF16 and FP8 checkpoints on Hugging Face and ModelScope alongside a technical blog, and keeping a date it had set when the model first appeared in September. The same day, the model was sitting second on OpenRouter's weekly token leaderboard — behind DeepSeek V4.1 Flash, ahead of Tencent's Hy4 preview — which is an awkward fact to pair with a download link, and that pairing is the story. What "open" buys you here is a license with commercial conditions attached, and a deployment that starts at a multi-node GPU cluster rather than a workstation.
What shipped, and when
StepFun announced the open-weights date when Step 5 Preview launched, months after the model leaked and the company said the weights would stay closed — StepFun's Step 5 Preview leaks at 600B — and the weights stay closed. The release missed the early hours of its own date: as of 01:55 UTC the repository still returned a 404, a gap Forkast turned into a piece about labs missing announced dates, before the checkpoints went live later in the day. Both the company's own post and the model cards are up now, and Chinese tech outlet 智东西 Zhidx and Pandaily independently reported the same arrival — the weights on Hugging Face and ModelScope, plus a technical blog explaining how the model was built.
The architecture is a memory argument
Step 5 Preview is a sparse mixture-of-experts model: 600 billion total parameters, with 27 billion active per token, and a context window of about one million tokens that takes text, images and video. The design choice worth understanding is depth. StepFun stacks 92 Transformer blocks — as many as trillion-parameter open models — on a narrower base, on the argument that long agentic tasks get little time for explicit chains of thought and lean on implicit multi-hop reasoning while reading context, which needs layers to propagate through. Deep, narrow stacks are cheaper than wide ones at comparable reasoning depth, which is how the model lands where it does while activating 4.5 percent of itself.
Long context is the second half of the same argument. The attention layers interleave in a 3:1 ratio: three quarters are gated sliding-window attention, and the remaining quarter is sparse full attention using what StepFun calls sparse GQA — grouped-query attention with a sparse index on top. The model selects 4,096 tokens out of the full history, in blocks of eight, rather than attending to everything. StepFun reports that the coarse selection covers 78.06 percent of what a finer-grained reference scheme covers at 79.09 percent, meaning most of the compute goes on evidence that does not change the answer. Training stability gets the same treatment: hidden states and normalisation scale factors are held in FP32 inside an FP8 training run, and only about 0.00004 percent of gradients are clipped per layer to suppress spikes. The benchmark numbers are the company's own — an Artificial Analysis Intelligence Index score of 44, DeepSWE v1.1 at 67.7, just ahead of Kimi K3 Max at 67.5 and GLM-5.3 Max at 66.9 while trailing GPT-6 Astra at 74.1 and Claude Opus 5 at 74.0.
Who can actually run it
Start with arithmetic. Six hundred billion parameters in BF16 is roughly 1.2 TB of weight data before a single token of KV cache — and with a one-million-token context, the cache is not the rounding error. The model card itself describes a 16-shard checkpoint and a two-node launch, and states plainly that the full BF16 conversion workflow has not yet been validated end to end. The FP8 checkpoints halve the footprint to about 600 GB, which is still four 80 GB accelerators before you serve anything. There is no quantised GGUF-style path announced, and none of this is a scandal — it is what hosting a frontier-scale sparse model costs. It is also why the 15.1 trillion tokens Step 5 Preview moved across OpenRouter in the week through October 13 were not served by hobbyists — that total put it second on the weekly leaderboard, well behind DeepSeek V4.1 Flash at 40.5 trillion and ahead of Tencent's Hy4 preview, the Chinese model that took the top spot as Chinese labs led for 19 weeks. OpenRouter routes the model to two providers, DeepInfra and StepFun itself, and that is where the traffic runs.
The license is the second gate. The weights ship under the StepFun Model License 1.0, which the company describes as Apache-2.0-style with additional attribution and commercial-license conditions. That is not an OSI-approved open-source license, and it is a different posture from the permissive releases developers compare Chinese labs against — the distinction between "you can download it" and "you can build a business on it" that we walked through in How to — check a model's license before you ship on it. For most teams the practical effect is unchanged: they were going to call the API anyway, at $1 per million input tokens, $2.70 per million output tokens and $0.05 per million cached reads.
The distribution deals fill in the rest of the picture. StepFun put Step 5 Preview inside Tencent's WorkBuddy for a free run through October 25 — free at the point of use, on someone else's infrastructure. The open weights are a credibility signal, a hiring signal and a hedge against Western API access; they are not the business. The business is tokens, and the same lab is now selling tokens bundled with hardware in China, where it just opened pre-orders on an agent-first phone with a twelve-month usage plan attached. Nothing of that shape exists for Step 5 in the West.
What the skeptics say
Three things. First, every headline number here is self-reported; the one independent measurement available in the model's own release materials is the deployment geometry, and that geometry says "cluster." Second, the usage lead is narrower than "number two globally" suggests: DeepSeek V4.1 Flash moved 40.5 trillion tokens in the same week, and OpenRouter warns that providers count tokens differently, so the ranking is directional. Third, the striking "one-eighth of Claude Opus 5's cost per task" figure is a cost-per-task average that depends entirely on the mix of tasks and on the token counts each provider reports.
The fair criticism is not that Step 5 Preview is overrated. It is that "open weights" is doing promotional work it has not earned here — the weights are open to anyone with a data centre, a license review, and a reason to run a model that two hosted providers will serve for the price of a coffee. That is a real option, not a false one. It is just not the community-scale openness the phrase implies.
What to watch
- Whether any independent host publishes throughput and quality numbers at the full one-million-token context, rather than short-context smoke tests.
- Whether the BF16 conversion path gets validated end to end, or FP8 becomes the only practical checkpoint.
- Whether StepFun clarifies what the commercial conditions in its license actually require — the answer decides which startups can ship on this model.
- Whether DeepSeek's next release answers the price-per-task claim directly.
If "open weights" can mean a bespoke license and a two-node cluster, what would you want a lab to disclose before the phrase is allowed in its marketing? Tell us in the comments.
