Infrastructure partners vLLM and SGLang launched Day 0 support for the model upon its release. vLLM added compatibility for Blackwell and Hopper GPUs and introduced RL training through verl and an NVFP4 version of the weights. SGLang implemented MTP decoding and cache aware routing, which the project noted is already serving the model in production environments.
Tuesday, October 20, 2026
Story
StepFun Opens 600B Parameter Step 5 Preview as Open Weights Flagship
AI6 days ago
The company's latest AI architecture is now available for download on Hugging Face and ModelScope in BF16 and FP8 precision. This flagship model, named Step 5 Preview, utilizes a sparse Mixture of Experts (MoE) architecture with 600B total parameters and 27B active per token. It features a 1M token context window and native vision capabilities, designed specifically for agentic workflows such as software engineering, coding, and professional knowledge work in finance.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.