StepFun Releases Step 5 Preview Open Weights: 600B MoE, 27B Active, 1M Context

StepFun has put out Step 5 Preview as open weights, with FP8 and BF16 checkpoints now live on Hugging Face and ModelScope. The company is pitching it as a model for long, messy jobs: coding agents, document analysis, and finance workflows that keep going for a while instead of answering one neat prompt and stopping.

The numbers are the headline. Step 5 Preview has 600B total parameters, but only 27B activate per token. It also stretches to 1M tokens of context, which is the kind of window that makes large codebases, long documents, screenshots, charts, and even video less awkward to work with. StepFun says the model is aimed at software engineering and professional knowledge work, not casual desktop tinkering.

This is not a “run it on your laptop” release. StepFun lists FP8 deployment at 4 B300 GPUs or 8 H200 GPUs, and BF16 at 8 B300s or 16 H200s. The FP8 checkpoint alone is 650.32 GB. The company also says Transformers support is offline generation only, and the full BF16 shard workflow is not yet validated end to end.

On performance, StepFun’s own chart puts Step 5 Preview near open-weight peers GLM-5.3 and Kimi K3 on coding, while trailing GPT-6 Astra and Claude Opus 5. Independent numbers from Artificial Analysis give it 88.3% on AA-LCR long-context reasoning and 43.7 on the Intelligence Index. But the weak spots are obvious too: 33.3% on Terminal-Bench 4.0, 0.4% on SRE Bench, and 9.2% on Harvey’s Legal Agent Benchmark.

The model uses a sparse Mixture-of-Experts design with 92 transformer blocks, gated sliding-window attention mixed with full attention, and an MTP-3 head that StepFun says gives about a 30% end-to-end speedup. The API launched on September 20, and the weights followed on October 14. Pricing starts at $1.00 per 1M input tokens and $2.70 per 1M output tokens, with cache hits at $0.10 per 1M.