Editorial illustration for StepFun Releases Step 5 Preview Open Weights: 600B MoE, 27B Active, 1M Context
StepFun Releases Step 5 Preview Open Weights: 600B MoE,...
StepFun put its flagship model out in the open on October 14, dropping FP8 and BF16 checkpoints of Step 5 Preview on Hugging Face and ModelScope. The API version launched a few weeks earlier, on September 20, so the weights release closes the gap for anyone who wants to run this thing themselves instead of calling StepFun's servers.
Step 5 Preview is a sparse Mixture-of-Experts model, 600 billion parameters total, with only 27 billion active on any given token. It handles up to 1 million tokens of context, which StepFun is pitching at long-horizon agentic work: coding agents that need to track multi-file changes across a sprawling codebase, document analysis pipelines chewing through charts and screenshots, finance workflows that need a model to hold a lot of state at once.
That scale comes at a real hardware cost, and the model's benchmark results land it somewhere between the top closed models and the best open-weight competitors depending on the task. StepFun's own engineering team has detailed how the architecture is built to make that 27B-of-600B split work in practice.
StepFun has released Step 5 Preview as open weights. FP8 and BF16 checkpoints are live on Hugging Face and ModelScope. Step 5 Preview is a sparse Mixture-of-Experts model from StepFun built for long-horizon agentic work.
Why this matters
StepFun just handed developers a 600B-parameter model that only costs 27B active parameters to run per token, with a 1M-token context window, and put the weights on Hugging Face and ModelScope for anyone to download. That's a meaningfully different proposition than another closed API. For teams building coding agents, document pipelines or finance tools, self-hosting at this scale means you control the data, the latency and the cost curve, instead of renting it by the token from a lab that can change pricing or access overnight.
The catch is obvious: running 600B total parameters, even sparsely activated, still demands serious infrastructure, and "open weights" doesn't mean "easy to deploy." We'd want to see independent benchmarks on the multi-file coding and long-document claims before treating this as a drop-in replacement for proprietary agents. But the trend is worth tracking. If Chinese labs keep shipping frontier-scale MoE models as open weights, the gap between what you can self-host and what you have to rent keeps narrowing, and that changes who gets to build what.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Box CEO Aaron Levie Says AI Will Drive "Biggest Cybersecurity Upgrade
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives Perplexity Open-Sources SPACE Firecracker MicroVM Sandbox for AI Agents
From Archives Nvidia Says Nemotron 3.5 Super VL Outperforms Gemini Ultra in Agent Tests
Further Reading
- StepFun Releases Open Weights for Step 5 Preview, Publishes Technical Blog on Its Architecture - Pandaily
- StepFun launches Step 5 Preview with 600B parameters, 1M context and open weights coming October 15 - Data Studios
- Step 5 Preview: Advancing the Pareto Frontier - StepFun
- Step 5 Preview Model Card - Hugging Face
