StepFun Releases Step-5 Preview Open Weights: 600B Parameters on Hugging Face

What happened

On October 15, StepFun published Step-5 Preview weights on Hugging Face. The same day, NVIDIA said its NVFP4-quantized build runs at nearly 8,000 tokens per second per GPU on the GB300 platform, calling it StepFun's largest open-weights release to date.

Step-5 Preview targets agentic work: per the company, it is built for software engineering, professional knowledge work and finance-style research, with long-horizon tool use; when the API version launched on September 20, StepFun promised weights on October 15.

Key facts

1. Specs: Sparse MoE: about 600B total parameters, about 27B active per token (roughly 4.5%); 1M-token context; text, image and video input, text output.
2. API pricing: $1.00 per million input tokens and $2.70 per million output tokens (reasoning tokens count as output); Artificial Analysis puts the medians for comparable models at $1.88 input and $10.00 output.
3. Independent score: Artificial Analysis Intelligence Index: 44, vs a median of 24 for similarly priced reasoning models; measured API output speed about 99.8 tokens per second.
4. Company-reported scores: FrontierFinance 66.4 (Claude Opus 5: 69.7, GPT-6 Astra: 55); DeepSWE v1.1 67.7; StepFun also ran two 24-hour agent experiments: tuning an H100 kernel to 508 TFLOPS (Claude Opus 5: 493), and lifting Qwen3-30B-A3B on AIME24 from 53.3% to 60% via automated post-training.
5. NVIDIA deployment: NVFP4 build on GB300; NVIDIA claims nearly 8,000 tokens per second per GPU.

Background

Step-5 Preview is StepFun's flagship model, launched API-first on September 20 with weights promised for October 15. By simple arithmetic, 600B parameters need about 1.2TB of memory in BF16 before KV cache — self-hosting means multi-GPU server hardware.