AI Industry

StepFun Opens Step 5 Preview Weights: 600B, 1M Context

Oct 15, 2026·4 min read·By Rahish Saifi

Photo of an unlocked antique brass padlock and key on sage-green linen, representing StepFun releasing Step 5 Preview open weights

StepFun published the open weights of Step 5 Preview on October 15, 2026, delivering on the release timeline it set when the model launched on OpenRouter and partner tools a week earlier. Checkpoints are on Hugging Face—including FP8 (stepfun-ai/Step-5-Preview), BF16, and NVFP4 variants—and on ModelScope, with the primary FP8 repository updated Oct 15, 2026 per Hugging Face metadata.

The model card describes Step 5 Preview as StepFun’s flagship model for agentic work, with frontier-level performance across software engineering and professional knowledge work and particular strength in finance. It is a sparse mixture-of-experts (MoE) design with 600B total parameters and 27B active parameters, supporting up to 1 million tokens of context and accepting text, images, and video for sustained reasoning and tool use.

What the weights include

StepFun shipped three precision flavors on Hugging Face: the default FP8 checkpoint, a BF16 build, and an NVFP4 build. The card’s specification table lists a 600B parameter total, 27B active, context up to 1M tokens, text and image input, and text output; the introductory copy also notes video input for long-horizon agent workflows.

Capabilities highlighted on the card mirror the hosted product: repository-scale software engineering, multi-step tool calling with verification, multimodal analysis of documents and charts, and finance-oriented reasoning. Recommended inference settings on the card are temperature=1.0, top_p=0.95, and top_k=0, with published benchmark rows using the High reasoning setting.

Hosted API vs self-hosting

Developers who do not run their own GPUs can still call Step 5 Preview through StepFun’s platform. Official API documentation lists model ID step-5-preview, a 1M-token context window, text/image/video input, text output, and a 64k-token maximum output on the API. The docs support streaming, tool calling, JSON Mode and JSON Schema, prompt caching, and reasoning effort levels low, medium, and high.

The model first appeared on OpenRouter as stepfun/step-5-preview on October 8, 2026. OpenRouter quotes $1.00 per million input tokens and $2.70 per million output tokens, with a 1,048,576-token context window and up to 64,000 completion tokens. StepFun’s Oct 8 post on X promoted a week of free access through OpenCode, Cline, Nous Portal, and Kilo Code.

Deployment footprint

Self-hosting is realistic only at datacenter scale. The model card’s deployment table sizes FP8 serving on one node with 4× B300 or 8× H200 GPUs under vLLM or SGLang (OpenAI-compatible APIs), or offline generation with Transformers. BF16 needs 8× B300 on one node or two 8× H200 nodes (16 GPUs). StepFun documents tensor, expert, and combined parallelism modes, reasoning parsers, tool-call parsers, and speculative decoding flags for vLLM and SGLang.

Precision	Framework	Minimum hardware (per model card)	Interface
FP8	vLLM / SGLang	1 node: 4× B300 or 8× H200	OpenAI-compatible API
BF16	vLLM	1× 8× B300 or 2× 8× H200 (16 GPUs)	OpenAI-compatible API
FP8 / BF16	Transformers	8× H200 (FP8) or 16× H200 (BF16)	Offline generation

The card warns that a 1M-token limit covers combined input and output KV cache pressure; operators may need to lower --max-model-len or concurrency if memory is tight.

License and commercial use

Weights ship under the StepFun Model License 1.0, which StepFun describes as Apache-2.0-style with extra conditions. The LICENSE.txt in the Hugging Face repo governs use, modification, and distribution.

Commercial products or services with more than 100 million active users, or more than US$20 million in revenue, in any calendar month must display “Powered by StepFun AI” in a user-visible way. A separate commercial license is required for hardware makers that ship the model as a device feature (phones, vehicles, glasses, robots, smart-home products) and for any commercial use while operating a Model-as-a-Service business, if combined revenue from all sources (including affiliates) exceeds US$20 million over any consecutive 12-month period; internal use that exposes nothing to third parties is exempt. Models trained materially on StepFun model outputs count as derivatives under Section 1 of LICENSE.txt.

Third-party benchmarks (attributed)

Before weights landed, integrators published early scores on their own harnesses—useful signals, not StepFun audits.

Nous Research reported Step 5 Preview at 33.89 on its Hermes Index, matching GPT-6 Luna on that leaderboard. Cline said the model scores ahead of Kimi K3 and GLM-5.3 on DeepSWE in its testing. OpenRouter’s listing cites an Artificial Analysis Intelligence Index score of 43.7 for the hosted slug; that figure reflects API routing, not necessarily your local vLLM build.

What it means for developers and companies

Open weights chiefly help teams with GPU estates and compliance reasons to keep inference in-house. The FP8 path still assumes four Blackwell-class or eight Hopper-class GPUs before you count KV cache headroom for million-token sessions—far beyond a single developer workstation. Most product teams will keep paying per token on StepFun or OpenRouter while experimenting with weights in research sandboxes.

The license is permissive for startups but not frictionless for hyperscalers or device OEMs: attribution at nine-figure user scale and separate commercial terms for on-device or high-revenue MaaS still apply. Legal review before fine-tuning or redistributing derivatives is prudent, especially if training data includes model-generated outputs.

Step 5 Preview also underscores competition among Chinese labs shipping agentic MoEs with million-token windows alongside Western closed APIs. For finance and coding agents, the combination of long context, tool calling, and multimodal input is the selling point; for operators, the bottleneck remains capital expenditure on inference silicon, not tokenizer downloads.

What to watch

Track whether community fine-tunes appear under the derivative rules, and whether StepFun publishes NVFP4 serving recipes beyond the checkpoint drop. If you rely on OpenRouter, monitor provider uptime and cache pricing ($0.05/M cache read on OpenRouter’s listing) against StepFun’s direct API rate limits.