StepFun's Step 5 Preview: a 600B MoE model for agents, weights due 15 October

StepFun launched Step 5 Preview, a sparse Mixture-of-Experts model with a 1M-token context aimed at agent workloads. Open weights are promised for 15 October, not out yet.

In this article

Architecture and scale
Benchmarks and pricing
Open weights are not here yet

StepFun has launched Step 5 Preview, a sparse Mixture-of-Experts model aimed at long-horizon agent work such as coding, financial analysis and professional knowledge tasks. The model is available now through StepFun's API; the company says open weights will follow on 15 October.

Rows of black supercomputer cabinets with green and blue indicator lights in a data hall. — Photo: NASA Goddard Space Flight Center from Greenbelt, MD, USA (Public domain), Wikimedia Commons.

Architecture and scale

StepFun describes Step 5 Preview as roughly 600 billion total parameters with about 27 billion active per token, built on a 92-layer narrow-deep transformer rather than a wider network, which the company argues gives longer reasoning paths during long-context processing. It supports a one-million-token context window and takes both text and image input. Training reportedly relied on on-policy long-horizon reinforcement learning with techniques such as speculative decoding and FP8 MoE compute to speed up the process -- details that come from StepFun's own release material.

Benchmarks and pricing

According to Artificial Analysis, Step 5 Preview scores around 44 on its Intelligence Index, roughly matching Kimi K3 (max) and about one point behind GLM-5.3 and Qwen3.8 Max. The same benchmark source has it performing well on Humanity's Last Exam and CritPt physics questions, but trailing peers on agentic benchmarks such as GDPval-AA and Terminal-Bench 4.0, with a reported 43% hallucination rate. StepFun's listed API pricing is about $1 per million input tokens and $2.70 per million output tokens, with roughly 100 tokens per second of output -- notably cheap relative to comparable frontier-adjacent models, though these are list prices rather than audited unit economics.

Open weights are not here yet

As of publication, StepFun's Hugging Face repository for the model contains only a placeholder file, with no weights, license or model card. The company has committed to an open release on 15 October; once available, self-hosting the full model will need roughly 1.2 terabytes of memory in BF16 before accounting for KV cache, meaning multi-GPU server hardware rather than a single workstation.

Treat the benchmark and cost figures above as vendor- and third-party-reported until independent testing confirms them, and treat the 15 October open-weight date as StepFun's stated plan rather than a guaranteed release.

Sources: MarkTechPost, Pandaily.