StepFun Step-5-Preview — open BF16 weights released
In today's Spotlight · Picked by 3 of 4 modelsStepFun released the full downloadable files of its large Step-5-Preview AI model, which accepts text, images and video.Like a restaurant publishing its secret recipe for anyone to cook.
StepFun published the full BF16 weights for its flagship Step-5-Preview agentic model (600B-parameter sparse MoE, 27B active per token, 1M-token context, text/image/video input) on Hugging Face overnight.
Why it matters
First 600B-class Chinese open-weight model scoring in the global top-3 on Artificial Analysis (index 44) while costing ~1/8th of Claude Opus 5 per task; built natively for long-horizon agentic coding and tool use.
What you could build with it
An indie dev-tools startup could self-host Step-5-Preview to power a low-cost long-horizon coding agent for repo-scale refactors: the 1M-token context fits entire codebases, tool calling is native, and inference runs on the user's own GPU box at a fraction of frontier-API prices.
Does it hold up?
Too early to judge self-hosting — weights landed hours ago. API track record is strong (#2 on OpenRouter by usage pre-release). Caveat: license is the StepFun Community License, not Apache/MIT, so check commercial terms before productizing.
Built with StepFun Step-5-Preview
- Volanea: Step 5 Preview coding-agent review — 83.75% on KingBench 3article · Independent hands-on review ran eight agentic coding tasks (archery game, local Gemma fine-tuning workflow) scoring 67/80, above GLM 5.3 Flash and MiMo v2.6.
- Pollinations adds Step 5 Preview via OpenRouter with Vercel fallbackgithub · Pollinations wired the model into its free text API with OpenRouter primary and hidden Vercel fallback (text+image input).
- Vercel AI Gateway listing + free week on Cline, Nous, Kiloarticle · Model is callable as stepfun/step-5-preview on Vercel AI Gateway; OrcaRouter documents a free-access week on coding agents ending Oct 16.
- pollinations/pollinationsgithub · 5,198 ★ · Added Step 5 Preview via OpenRouter with Vercel fallback into its generative-AI API registry.
- q-xuan/agent-horizongithub · 1 ★ · Switched a GitHub Actions daily digest pipeline to the StepFun step-5-preview endpoint.
- miaai-lab/step-5-preview-100-html-filesgithub · 1 ★ · Gallery of 100 self-contained single-file HTML pages (each with prompt + screenshot) generated with Step 5.
Learn more
Official model card — architecture, benchmarks, deployment guide →
First spotted on hf: source.
More AI models
OpenAI publishes 722 AI-generated math manuscripts from an unreleased modelOpenAI released 722 mathematical manuscripts in 372 result families to GitHub under Apache-2.0, produced by an…models · JEV 0.74ElevenLabs Eleven v4 + v4 Turbo: new emotive TTS architecture with a 100ms real-time variantElevenLabs released Eleven v4 and Eleven v4 Turbo on September 28: a new text-to-speech architecture with inline…models · JEV 0.72Mistral Large 4 (Le Chonk): 1.05T-param open-weight MoEMistral's new flagship: a 1.05T-parameter mixture-of-experts multimodal model with 49B active params and 1M-token…models · JEV 0.72VoiceStudio: open-source, fully-local ElevenLabs alternative with voice cloning, dubbing and dictation in 646 languagesVoiceStudio is an open-source, fully-local voice AI studio: voice cloning from a 3-second sample, voice design, video…models · JEV 0.71JetBrains Mellum 2.1: 12B MoE coding model trained for agentsJetBrains released Mellum 2.1, a 12B MoE model (2.5B active parameters, 128K context) under Apache 2.0, trained mainly…models · JEV 0.69DeepSeek V4-Flash official API: public beta with upgraded agent capabilities, Responses API and Codex supportDeepSeek's official V4-Flash is now a public beta API with massively upgraded agent capabilities - benchmark scores…models · JEV 0.69
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.