StepFun has released open weights for its Step 5 Preview model, making FP8 and BF16 checkpoints available on Hugging Face and ModelScope. The sparse Mixture-of-Experts architecture features 600 billion total parameters with 27 billion active per token, alongside a 1 million token context window. Originally launching its API in September, the open-weights release targets developers building self-hosted coding agents, document processing pipelines, and complex financial workflows.
The model is built to navigate large codebases, execute multi-file changes, and analyze multi-modal assets like charts and videos. On benchmarks, Step 5 Preview scores 43.7 on the Intelligence Index, 88.3% on AA-LCR, and 66.4% on FrontierFinance. However, evaluations show performance gaps in specific domain tasks, such as scoring 0.4% on SRE Bench and 9.2% on Harvey’s Legal Agent Benchmark.
Deploying the model requires significant data center infrastructure, as the FP8 checkpoint requires 650.32 GB of storage. StepFun provides dedicated Docker images for vLLM and SGLang featuring OpenAI-compatible endpoints, while API access remains available starting at $1.00 per 1M input tokens and $2.70 per 1M output tokens.
Why it matters
Provides developers a self-hostable 600B MoE model offering long-context window capability for complex software engineering and agentic workflows.
Requires heavy data center infrastructure, with the FP8 checkpoint alone exceeding 650 GB of storage.
Exhibits domain-specific performance trade-offs, scoring well in general finance benchmarks but struggling on specialized legal and site reliability engineering tasks.
Source: marktechpost.com



