Step 5 Preview's full weights were officially opened for download by StepFun on October 15, 2026, in BF16 and FP8 versions, now listed on Hugging Face and ModelScope alongside a technical blog post, companion development tools and a deployment guide. The model itself debuted on September 20 with long-horizon agent work as its headline use case; open weights mean teams can finally run it inside their own infrastructure instead of calling it only through an API.
A sparse design aimed at the electricity bill of flagship capability
Step 5 Preview uses a sparse MoE architecture with 600B total parameters, only about 27B active per token, a one-million-token context window and native text and vision input. Several details in the technical write-up stand out: a 92-layer narrow-and-deep layout, Sparse GQA, training-stability work, and context compression folded into end-to-end policy training so an agent can tidy up its task state mid-run and keep going, backed by a million-scale verifiable environment pool and long-horizon reinforcement learning. The target scenes are explicit: AI coding, software engineering, professional knowledge work and finance, not casual chat. Trading activation ratio for cost is the same efficiency line as LFM2.5 running agents on a laptop with 24B total and only 2B active, except Step 5 pushes the scale into flagship territory.
A score of 44 and a 71-cent task: where it sits on value
On the Artificial Analysis Intelligence Index, Step 5 Preview scores 44, placing it in the global top three among open models. On the same basis, one index task costs it about $0.71, which the company calculates at roughly an eighth of Claude Opus 5, whose per-task total is about $5.86. API pricing is $1 per million input tokens and $2.7 per million output tokens. Developers are already voting with usage: OpenRouter data as of October 15 puts it second worldwide on both the weekly and daily token-usage charts. One caution: the 44 score and the cost comparison come from the benchmark provider and the vendor, and real gaps depend on task type, so long-horizon workloads deserve a sampled re-test on your own cases.
Once the weights land, do the deployment math first
On the serving side, the release supports multiple chip platforms and inference frameworks: vLLM or SGLang can host an OpenAI-compatible service, and Transformers covers offline inference, with hooks into your own data and tooling. But 600B total parameters set a floor that FP8 can lower without removing: multi-GPU serving and its operations cost remain. For smaller teams, the practical value of open weights is often not self-hosted inference at all, but fine-tuning, auditability and deep integration while data stays on-premises. Running your own long-task set to measure hit rate and per-task cost first, and only then choosing between self-hosting and the API, is the safer order.
