StepFun publishes open weights for Step 5 Preview, a 600B agent model

Chinese AI lab StepFun has released the weights for Step 5 Preview, its flagship model for agent work, on Hugging Face and ModelScope in FP8 and BF16 formats. MarkTechPost covered the release on October 16, 2026. The model first launched through StepFun's API on September 20.

Step 5 Preview is a mixture-of-experts model with 600 billion parameters, of which about 27 billion are used for each token. It accepts text and images, handles up to 1 million tokens of context, and is pitched at coding agents, long document analysis and finance work. Inside, it uses 92 layers that mix sliding-window and full attention, plus a multi-token prediction head that StepFun says speeds up output by about 30 percent.

On StepFun's own numbers it scores 67.7 on DeepSWE and 33.3 on Terminal-Bench 4.0, close to open rivals GLM-5.3 and Kimi K3 on coding but behind top closed models. Independent testing by Artificial Analysis gives it 43.7 on its Intelligence Index and a strong 88.3 percent on the AA-LCR long-context test. The license is not plain Apache: very large products must credit StepFun, and device makers need a separate commercial deal. Self-hosting needs at least four Nvidia B300 GPUs, while the API costs $1 per million input tokens and $2.70 per million output tokens.

Why it matters

Another frontier-scale open model gives teams a cheaper self-hosted option for long-context agents, but its license terms need checking before commercial use.

Source: MarkTechPost