Skip to content
Neural Daily

ModelsNews

StepFun Step 5 Preview: 600B open-weight agent model

China's StepFun released open weights for Step 5 Preview on October 15, its flagship model for agentic work. It is a sparse MoE with 600B total parameters, 27B active per token and a 1M-token context window.

By: Mikola Máté Patrik · · 2 min read

StepFun page: Step 5 Preview, Advancing the Pareto Frontier
Image: StepFun · Screenshot · Screenshot of official page · Source

What the model does

Step 5 Preview accepts text, images and video and targets long, tool-heavy agentic tasks: coding, software engineering and professional knowledge work, with particular strength in finance according to StepFun (StepFun). BF16 and FP8 weights are on Hugging Face and ModelScope, and it can be served as an OpenAI-compatible API with vLLM or SGLang (Hugging Face).

The technical blog describes a “narrow but deep” design: 92 transformer blocks that interleave Gated Sliding Window Attention with sparse full-attention layers (StepFun). StepFun says this gives agentic capability comparable to trillion-parameter models at much lower cost.

Benchmarks

StepFun benchmark chart: Step 5 Preview versus Kimi K3, GLM-5.3, GPT-6 Astra and Claude Opus 5 on eight tests
StepFun · Press kit · Vendor benchmark figure from the model card · Source

Source: StepFun. Step 5 Preview (High) versus Kimi K3, GLM-5.3, GPT-6 Astra and Claude Opus 5, all at Max settings.

In StepFun's own table, Step 5 Preview scores 67.7 on DeepSWE v1.1, 80.5 on ProgramBench and 66.4 on FrontierFinance. It beats the Chinese open models (Kimi K3, GLM-5.3) on most tests, but GPT-6 Astra and Claude Opus 5 lead on coding: on Terminal-Bench v4 they score 57.9 and 52.3 against Step 5 Preview's 33.3. On FrontierFinance it is ahead of GPT-6 Astra (55.0). These are vendor-reported numbers.

Licence

The model ships under the StepFun Model License 1.0, an Apache-2.0-style licence with additional attribution and commercial-use conditions. Read the LICENSE file before production use.

Why it matters

Alongside DeepSeek, Reflection and Meta, this is one of the week's big open-weight releases, tuned specifically for agent work. With 27B active parameters, serving costs are far lower than the 600B size suggests, although the full weights still need serious GPU memory. Worth testing for finance agents.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles