StepFun Releases Open Weights for Its 600B Step 5 Preview Model

Chinese AI lab StepFun has released the weights for its Step 5 Preview model as open source, joining a growing list of labs making large models available for self-hosting. The FP8 and BF16 checkpoints are live on Hugging Face and ModelScope as of October 14.

Step 5 Preview is a sparse Mixture-of-Experts model, meaning it holds 600 billion total parameters but only activates about 27 billion per token. It can process up to one million tokens of context and was built specifically for long-horizon agentic work like software engineering, document analysis and finance workflows.

The model first became available through the StepFun API on September 20. Open weights followed about three weeks later ¹.

How It Works

The model uses 92 transformer blocks in a narrow, deep design. It combines full attention layers with sliding-window attention in a 3-to-1 ratio and selects only 4,096 history tokens per query across blocks of 8 ¹. An MTP-3 speculative decoding head gives roughly a 30 percent end-to-end speedup. Training used FP8 mixed precision and reinforcement learning on million-level verifiable agentic environments.

Performance

On benchmarks, Step 5 Preview lands near open-weight rivals GLM-5.3 and Kimi K3 on coding tasks. StepFun's own DeepSWE v1.1 score was 67.7, versus 66.9 for GLM-5.3 and 67.5 for Kimi K3 ¹. However, it trails proprietary models like GPT-6 Astra and Claude Opus 5 .

Independent evaluations show the model scores 88.3 percent on Artificial Analysis's long-context reasoning test, its strongest result. Its weakest showing was 33.3 percent on Terminal-Bench 4.0, compared with 41.9 percent for GLM-5.3. Finance benchmarks were more promising, with 66.4 percent on FrontierFinance ¹.

Hardware and Cost

Running Step 5 Preview requires data-center GPUs. The FP8 checkpoint alone is over 650 gigabytes. The smallest listed setup is four B300 GPUs or eight H200 GPUs for FP8, or eight B300s or sixteen H200s for BF16 ¹. The BF16 shard workflow has not yet been fully validated end to end.

The API charges $1.00 per million input tokens and $2.70 per million output tokens, with cache hits at $0.10 per million ¹. Reasoning tokens are included in output pricing, and the API supports tool calling and JSON Schema output.

Developers interested in self-hosting can find the weights on Hugging Face, ModelScope and the StepFun GitHub repository.

Key facts

• Step 5 Preview has 600B total parameters but only 27B active per token
• Open weights are available in FP8 and BF16 on Hugging Face and ModelScope
• Requires at least 4 B300 or 8 H200 GPUs to run
• Scores 88.3% on long-context reasoning, 33.3% on Terminal-Bench 4.0
• API costs $1.00 per million input tokens and $2.70 per million output tokens

Sources

1. StepFun Releases Step 5 Preview Open Weights: 600B MoE, 27B Active, 1M Context — marktechpost.com
2. StepFun teases Step 5 Preview, a 600B/27B MoE, open weights coming Oct 15