StepFun Releases Step 5 Preview Open Weights: 600B MoE, 27B Activated, 1M Context
According to an October 16, 2026 MarkTechPost report, StepFun released Step 5 Preview as open weights: 600B total parameters, 27B activated per token, up to 1M context, with FP8 and BF16 checkpoints available on Hugging Face and ModelScope.
What changed
According to an October 16, 2026 MarkTechPost report, StepFun has released Step 5 Preview as open weights, with FP8 and BF16 checkpoints available on Hugging Face and ModelScope, while the API reportedly went live earlier on September 20 and the weights followed on October 14. The report describes the model as a sparse mixture-of-experts architecture with 600B total parameters, 27B activated per token, up to 1M tokens of context, and a narrow-and-deep design of 92 layers, with gated sliding-window attention and full attention interleaved at 3:1, full-attention layers selecting 4,096 historical tokens per query and scoring them in blocks of 8, and an MTP-3 speculative decoding head delivering roughly 30% end-to-end speedup. The report lists FP8 as requiring 4 B300 or 8 H200 GPUs, BF16 as requiring 8 B300 or 16 H200 GPUs, an FP8 checkpoint of about 650.32 GB, and vLLM and SGLang images with OpenAI-compatible interfaces. The report also gives StepFun's self-reported benchmarks alongside independent results from Artificial Analysis and Vals AI, and notes that in the self-evaluation opponents used Max effort while Step 5 Preview used High reasoning, meaning the comparison is not like-for-like.
What it means for developers
For developers, this release puts 1M context and long-horizon agentic capabilities into a self-hostable option, but the bar is data-center class: the minimum configuration is 4 B300 GPUs (FP8), and BF16 requires a single node of 8 B300 or a dual node of 16 H200, which makes it hard to fit into ordinary inference clusters or on-device deployment. On licensing, the report explicitly lists the StepFun Model License 1.0; open weights do not equal an open-source license, and commercial use, redistribution and derivative terms must be checked against the license text itself, since visible weights do not imply the scope of granted rights. The benchmarks cited in the report show a clearly split capability profile: 88.3% on AA-LCR long-context reasoning and 66.4% on FrontierFinance, but only 33.3% on Terminal-Bench 4.0 (versus 41.9% for GLM-5.3), 0.4% on Vals AI's SRE Bench, and 9.2% on the Harvey legal agent benchmark, suggesting that long-context reading and terminal-style multi-step operation may be strengths of different magnitude. The report also says the API supports low, medium and high reasoning effort, tool calling and JSON Schema output, which is relatively low retrofit cost for teams plugging it into an existing OpenAI-compatible pipeline. Reported facts and inference need to be kept apart: the benchmarks above are relayed by the source, self-reported and independent results use different settings, and none were re-verified for this article.
Integration and evaluation
Choose based on the capability profile rather than a total score. First decision: if the workload is mainly long documents, charts or financial-report analysis, the 88.3% AA-LCR and 66.4% FrontierFinance figures in the report are signals worth reproducing, and self-hosting teams can start with the FP8 4x B300 configuration on vLLM or SGLang at the report's recommended temperature 1.0, top_p 0.95 and top_k 0 for a small-sample comparison, provided they already have B300 or equivalent memory resources; the evaluation metric could be the structured-output field accuracy on their own long-document extraction task, compared in pairs against the current baseline. Second decision: if the workload is terminal operation, multi-step command execution or SRE-style automation, the 33.3% Terminal-Bench 4.0 and 0.4% SRE Bench figures in the report suggest doing a failure-mode analysis before deciding whether to adopt, given that the task has low tolerance for wrong commands; the evaluation metric could be end-to-end task success rate and number of human interventions on a fixed step set, with an adoption threshold. In addition, teams evaluating the API route should check current billing on StepFun's official pricing page instead, since API and cache prices can change. On integration cost, OpenAI-compatible images and JSON Schema support can reduce the retrofit effort, but before going live you should verify the stability of tool calling and structured output under your own schema.
Limits and open questions
Step 5 Preview does not currently appear in KimiSeek's live catalog, and the API availability described in the report refers to StepFun's official channel, not to this site having listed that version; the presence of other versions in the same family or with similar names in the catalog does not mean this version is available. The benchmarks, hardware configurations and prices described here are all cited from the October 16, 2026 MarkTechPost report, in which StepFun's self-reported data and its opponents did not use equal reasoning effort, and the independent results are relayed from Artificial Analysis and Vals AI; none were re-tested for this article and should not be treated as verified conclusions. The report explicitly labels it a Preview, so the architecture and evaluation settings may change in later versions; Transformers only supports offline generation, and the full BF16 sharding workflow has not yet been verified end to end, which is another unknown in self-hosting evaluation. The license is the StepFun Model License 1.0, and specific commercial and redistribution terms must be checked in the original text. Prices and regional availability may change, so refer to the official pricing and documentation pages.