StepFun opens Step 5 Preview weights: 600B MoE, 27B active, 1M context

StepFun has released Step 5 Preview as open weights: a 600-billion-parameter Mixture-of-Experts model with 27B active parameters per token and a context window of up to 1 million tokens. FP8 and BF16 checkpoints went live on Hugging Face and ModelScope on October 14, roughly three weeks after the model's API launched on September 20.

The release matters because self-hostable models at this scale are rare. StepFun positions Step 5 Preview as a flagship for agentic work — software engineering, professional knowledge work and finance — and the open weights give developers a new option they can run on their own hardware for coding agents, document analysis and finance workflows.

The catch is the hardware bill. This is a data-center model, not something you boot on a workstation.

What is Step 5 Preview?

Step 5 Preview is StepFun's flagship model for agentic work. The model navigates large codebases, coordinates multi-file changes, runs code and iterates on results. It also analyzes long documents, charts, screenshots and video to produce structured reports.

Text and image input ship with the weights; video arrives through the API. The model is licensed under the StepFun Model License 1.0.

In short: it's built for long-horizon tasks, not chat.

How does the architecture work?

Only 27B of the 600B parameters activate per token. That's the Mixture-of-Experts design — the model routes each token to a small slice of its total capacity, which keeps inference costs closer to a mid-size model while retaining the knowledge of a large one.

StepFun's tech team published the internals:

• Depth: 92 transformer blocks in a narrow, deep design.
• Attention mix: gated sliding-window attention and full attention layers, interleaved 3:1.
• Sparse GQA: full-attention layers pick 4,096 history tokens per query, scored in blocks of 8.
• Speculative decoding: an MTP-3 head gives roughly 30% end-to-end speedup.
• Training: FP8 mixed precision, with RL on million-level verifiable agentic environments.

The sparse attention scheme is what makes the 1M-token context practical — instead of attending to everything, each query samples a fixed slice of history.

How does it perform on benchmarks?

StepFun ran its own evaluations at the High reasoning setting, while rivals in its comparison chart ran at Max effort — so the numbers are not like-for-like.

Results reported on the Hugging Face model card:

• DeepSWE v1.1: 67.7 (SWE-agent harness, temperature 1.0, top_p 0.95)
• Terminal-Bench 4.0: 33.3
• Agents' Last Exam (ALE-CLI): 29.5

Independent numbers from Artificial Analysis via OpenRouter paint a fuller picture. The model scores 43.7 on the Intelligence Index, hits 88.3% on AA-LCR long-context reasoning and 46.5% on Humanity's Last Exam, and lands at 54.3% on GDPval-AA.

The 88.3% AA-LCR score is the standout — long-context reasoning is exactly what the architecture was built for.

The weak spots are just as clear. Vals AI results show SRE Bench at 0.4% and Harvey's Legal Agent Benchmark at 9.2%. On Terminal-Bench 4.0, the model's 33.3% trails GLM-5.3 at 41.9%. On finance, StepFun's chart shows 66.4% on FrontierFinance — that beats GLM-5.3 and Kimi K3, but trails Claude Opus 5 at 69.7%.

Overall: near open-weight peers GLM-5.3 and Kimi K3 on coding, behind GPT-6 Astra and Claude Opus 5, exceptional at long-context reasoning.

How does it compare with other open-weight models?

Feature	Step 5 Preview	GLM-5.3	Kimi K3
Total parameters	600B	753B (HF file size)	2.8T
Active parameters	27B	Not disclosed	104B
Context window	1M	1M	1M
Input modalities	Text, image (API adds video)	Text only	Text, image
Weight formats	FP8, BF16	BF16, FP8	MXFP4
Minimum listed self-host setup	4x B300 (FP8)	Not disclosed	Not disclosed
DeepSWE v1.1 (StepFun-reported)	67.7	66.9	67.5
Terminal-Bench 4.0 (StepFun-reported)	33.3	41.9	12.6
FrontierFinance (StepFun-reported)	66.4	64.1	62.6
API price per 1M tokens (in / out)	$1.00 / $2.70	Not disclosed	$3.00 / $15.00

Benchmark rows come from StepFun's comparison chart, as transcribed by CellCog and eesel.

Step 5 Preview sits in a useful middle ground: much smaller than Kimi K3's 2.8T total parameters, with the lowest active parameter count and the cheapest listed API pricing of the three.

What hardware does it need to run?

The FP8 checkpoint alone is 650.32 GB. Listed deployment configurations:

• FP8 on vLLM or SGLang: 4 B300 GPUs or 8 H200 GPUs, each on 1 node.
• BF16 on vLLM: 1 node with 8 B300 GPUs, or 2 nodes with 16 H200 GPUs.
• Transformers: offline generation only. The full BF16 shard workflow is not yet validated end to end.

vLLM and SGLang ship dedicated Docker images with OpenAI-compatible APIs. On Blackwell hardware, expert-parallel modes use the MegaMoE backend. StepFun recommends temperature 1.0, top_p 0.95 and top_k 0.

The takeaway: the FP8 route on four B300s is the cheapest door in, and it's still a serious cluster.

How much does the API cost?

For teams that skip self-hosting, the StepFun API charges $1.00 per 1M input tokens and $2.70 per 1M output tokens. Cache hits cost $0.10 per 1M. Output pricing includes reasoning tokens.

The API supports low, medium and high reasoning effort, tool calling and JSON Schema output. Against Kimi K3's $3.00 in / $15.00 out, Step 5 Preview undercuts the closest open-weight rival by several multiples on output price.

Why this release matters

Open weights at 600B total parameters are still a minority of the frontier-adjacent market, and most agentic deployments currently run on closed APIs. Step 5 Preview gives operators a self-hostable alternative with a 1M-token window, competitive coding scores and a clearly documented hardware path.

The benchmark gaps — 0.4% on SRE Bench, 33.3% on Terminal-Bench 4.0 — show where the model isn't ready yet. But with weights, a GitHub repo and API docs all public as of October 14, teams can now test those limits on their own infrastructure rather than take StepFun's chart at face value.