A Beijing lab opened the weights of a model most people had never heard of — and within days it was the second most-used model on OpenRouter, the aggregator developers actually pay to call. That is not a benchmark victory lap. It is a usage vote, and usage votes are harder to fake than leaderboard scores.
The model is Step 5 Preview from StepFun (阶跃星辰). Released on October 15, 2026, it arrived with the full open-source kit: BF16 and FP8 weight versions, a technical blog, deployment guides, and tooling, all mirrored on Hugging Face and ModelScope.
The specs that matter
Step 5 Preview is a sparse mixture-of-experts (混合专家, MoE) model with 600 billion total parameters and only 27 billion active per token. In plain terms: it is enormous to train but cheap to run, because each request wakes up a small slice of the network.
- Context (上下文): one million tokens, enough to hold a long document, a long chat, and many rounds of tool calls in a single pass.
- Modalities: native text and vision input — it is multimodal (多模态), not text-only with a bolted-on vision head.
- Target: long-horizon agent (智能体) tasks — the kind where a model has to plan, call tools, fix its own mistakes, and keep working for a long time.
The design leans into "do real work" rather than "chat well." StepFun says the model was built for software engineering, professional knowledge work, and finance.
Why it is built the way it is
The technical blog describes a deliberately unusual architecture:
- A 92-layer narrow-deep structure rather than the fatter, shallower stacks common in many MoE models.
- Sparse GQA (grouped-query attention) to keep the memory footprint of long contexts manageable.
- Training-stability optimizations and a million-scale verifiable environment used for long-horizon reinforcement learning.
- Most interestingly, StepFun folded context compression into the model's end-to-end policy training — so the agent can tidy up its own task state and pick the work back up, instead of forgetting where it was when the context gets long.
That last point is the actual thesis of the release. Long-horizon agents fail less because they are smarter and more because they lose the thread. If a model can compress and resume its own state, it survives longer jobs.
The OpenRouter #2 signal
According to OpenRouter data cited as of October 15, 2026, Step 5 Preview ranked second globally on both the weekly and daily token-usage charts. OpenRouter is the "water meter" of real developer traffic: the ranking reflects who is actually sending code, documents, and business data through a model, not who climbed a curated eval.
A few caveats worth holding in your head:
- Usage rank measures popularity and price-attractiveness, not capability. A cheap, capable-enough model will climb fast.
- Being second on OpenRouter is a strong adoption signal, but it is not the same as beating frontier closed models on hard benchmarks.
- The figure is a point-in-time snapshot; rankings move weekly as new models land.
Still, for an open-weight Chinese model to reach that position is a notable data point in the "open models are becoming the default" story.
How to actually run it
StepFun shipped deployment paths that avoid vendor lock-in:
- Serve it through vLLM or SGLang with an OpenAI-compatible API.
- Run offline inference with Transformers.
- It is adapted to multiple chip platforms and inference frameworks, so it is not tied to one accelerator vendor.
Because the weights are open, anyone can self-host, fine-tune, and wire it into their own data and tools — the property that makes "open-weight" (开放权重) different from "open API."
Trend reading
According to Li An, chief scientist at BrainNet (脑机网), "a model that reaches the top of a real usage chart the week its weights go public is the clearest signal yet that open Chinese models are competing on usefulness, not just on license generosity."
Honest limitations
- The 600B/27B parameter split and the 92-layer design are StepFun's published specs; we have not independently re-measured them, though the weights are downloadable for anyone to verify.
- "OpenRouter #2" is a usage-popularity metric reported via Chinese outlets citing OpenRouter; it is not an independent capability benchmark and is volatile week to week.
- StepFun's claims about context-compression training and long-horizon stability are descriptive, not yet backed by a published third-party evaluation we can cite.
- Running a 600B model — even at 27B active — still demands serious hardware; the "cheap to run" claim is relative to its size, not to a 7B model.
- FP8 weights trade some precision for smaller footprint; teams with strict numeric needs should validate before production.
What readers can do now
- Pull the FP8 weights and stand up a vLLM endpoint on hardware you already own to test the agent loop on a real long task.
- Compare your token spend on Step 5 Preview versus your current closed model on the same workflow before trusting the popularity ranking to mean "cheaper for you."
- If you depend on it in production, pin a specific weight commit and re-evaluate after the formal (non-Preview) release, since Preview checkpoints change.