--- title: StepFun's Step 5 Preview: 600B MoE model scores 67/80 on... description: StepFun released Step 5 Preview, a 600B-parameter mixture-of-experts model with 27B active parameters per token, supporting text and image input, a 1M-token... canonical: https://daily.dev/posts/stepfun-s-step-5-preview-600b-moe-model-scores-67-80-on-coding-tests-and-arrives-on-vercel-s-ai-gat-ndoie9yjg twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: StepFun's Step 5 Preview: 600B MoE model scores 67/80 on coding tests and arrives on Vercel's AI Gateway | daily.dev og:description: StepFun released Step 5 Preview, a 600B-parameter mixture-of-experts model with 27B active parameters per token, supporting text and image input, a 1M-token... og:url: https://daily.dev/posts/stepfun-s-step-5-preview-600b-moe-model-scores-67-80-on-coding-tests-and-arrives-on-vercel-s-ai-gat-ndoie9yjg og:image: https://api.daily.dev/og/posts/NDOiE9yjg.png og:image:alt: StepFun's Step 5 Preview: 600B MoE model scores 67/80 on coding tests and arrives on Vercel's AI Gateway og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # StepFun's Step 5 Preview: 600B MoE model scores 67/80 on coding tests and arrives on Vercel's AI Gateway **[Collections](https://daily.dev/sources/collections)** · 3 min read · 0 upvotes · 1 comments ## Summary StepFun released Step 5 Preview, a 600B-parameter mixture-of-experts model with 27B active parameters per token, supporting text and image input, a 1M-token context window, and up to 64K output tokens. Pricing is $1/M input tokens, $5/M cached input, $2.70/M output tokens, with full weights due October 15th. A sponsored hands-on review ran it through eight Kingbench 3 coding tasks inside the Open Code agent, scoring 67/80 (83.75%), beating GLM 5.3 Flash and MIMO v2.6 variants but trailing GLM 5.3. Strongest results were an archery game and a local fine-tuning workflow that trained and served a Gemma model, though its generated training data had factual errors. The model is now available on Vercel's AI Gateway as `stepfun/step-5-preview`, usable in coding agents like Claude Code, Codex, and Cursor via the Vercel CLI, targeting agentic coding, knowledge work, and financial analysis. ## Content StepFun's Step 5 Preview is a sparse mixture-of-experts model with 600B total parameters and 27B active per token. It reached #1 on OpenRouter Trending a day after launch. Open weights are due on October 15. ## Specs and pricing The context window is 1M tokens. One hands-on review lists 64K output tokens, while @omarsar0 says output can go up to 1M, so treat the output limit as unsettled. That review puts API pricing at $1 per million input tokens, $5 for cached input, and $2.70 per million output tokens. The model takes text and image input. StepFun is pitching it at real engineering work: fixing bugs in unfamiliar codebases, implementing features across several files, and refactoring so the changes run and pass checks. Frontend work is the other focus. Give it a screenshot, mockup, or PRD and it can return a working web page, or build a playable browser game from a set of assets. It also reads annual reports and operating data and turns them into financial reports, editable spreadsheets, and slide decks. ## Where you can use it - Kilo Code, Cline, Hermes Agent, and OpenCode support it, so you can keep your workflow and swap the model. - Vercel's AI Gateway offers it as `stepfun/step-5-preview` through the AI SDK. It can also be set up in Claude Code, Codex, and Cursor via the Vercel CLI. The gateway charges no markup on inference, including bring-your-own-key requests, and has usage reporting, zero data retention support, API key budgets, and routing rules. - vLLM had day-0 support. The @vllm_project team described work on sparse attention, MoE, multi-token prediction, and vision, with serving across Blackwell and Hopper GPUs. That covers the sparse indexer, FP8 MegaMoE, and cache transfer. ## Early impressions @omarsar0 has used it as a coding agent since early access. His take: it checks its own work and stops when the task is done, which makes it a good fit for long agent runs you don't watch closely. A sponsored review ran it through eight Kingbench 3 coding tasks inside the Open Code agent. It compared the model against MIMO v2.6 Pro and Flash and GLM 5.3 and GLM 5.3 Flash. The two best results: - A fully playable archery game with wind and moving targets scored 9.5/10. - A complete local fine-tuning workflow trained a Gemma model and served it through a local web app, also 9.5/10. The generated training data did contain factual errors. The other tasks were an elevator simulation, a 3D contact lens case, a folding table, a panda SVG, and a wristwatch. They scored between 6 and 8. The total was 67/80 (83.75%). That beats GLM 5.3 Flash and both MIMO reference scores, but trails GLM 5.3. Since the review was sponsored, the scores deserve some skepticism. Still, the agent-style behavior @omarsar0 describes matches what the benchmark tasks reward, and the open weights should let others check the claims soon. ## Questions this post answers ### What is StepFun's Step 5 Preview model and what are its specs? Step 5 Preview is a 600-billion-parameter mixture-of-experts model from StepFun with 27 billion parameters active per token. It accepts text and image input, offers a 1-million-token context window, and can output up to 64K tokens. API pricing is $1 per million input tokens, $5 per million cached input tokens, and $2.70 per million output tokens. Full weights are planned for release on October 15th. _daily.dev tracks new model releases like this for developers weighing coding-agent options._ ### How did StepFun Step 5 Preview score on coding benchmarks compared to GLM 5.3? Step 5 Preview scored 67 out of 80 (83.75%) across eight Kingbench 3 coding tasks run inside the Open Code agent, placing it above GLM 5.3 Flash and both MIMO v2.6 Pro and Flash scores, but behind GLM 5.3. Its strongest results were a fully playable archery game and a local fine-tuning workflow that trained and served a Gemma model, both scoring 9.5/10. _Developers comparing coding models can follow benchmark results like these on daily.dev._ ### How do I use StepFun's Step 5 Preview model in Claude Code or Cursor? Step 5 Preview is available through Vercel's AI Gateway under the model name stepfun/step-5-preview and can be called from the AI SDK. It can also be configured in coding agents such as Claude Code, Codex, and Cursor through the Vercel CLI. AI Gateway adds no markup to inference pricing, including Bring Your Own Key requests, and offers usage reporting and API key budgets. _daily.dev helps developers setting up new models in their coding agents stay current on options._ ## Community discussion Top comments from developers on daily.dev. **@bickov** · 0 upvotes > Does the 1M window change anything for screenshots, or does Step 5 resize images before they count? On Claude 4.7+ the long edge gets cut to 2576 px, so a 2540x18410 page capture reads as 355x2576 and 16 px text ends up about 2.2 px tall. A bigger context does not fix that. What is the image size cap here? ## Similar posts on daily.dev - [Step 3.5 Flash is Free in Kilo](https://daily.dev/posts/step-3-5-flash-is-free-in-kilo-hvrjnrdqi) · Kilo Blog · 2 upvotes · 0 comments - [GLM-5.2: Built for Long-Horizon Tasks](https://daily.dev/posts/glm-5-2-built-for-long-horizon-tasks-stmnmwjok) · Hugging Face · 33 upvotes · 4 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#ai-coding](https://daily.dev/tags/ai-coding), [#mixture-of-experts](https://daily.dev/tags/mixture-of-experts), [#ai-gateway](https://daily.dev/tags/ai-gateway) [View this post on daily.dev](https://daily.dev/posts/stepfun-s-step-5-preview-600b-moe-model-scores-67-80-on-coding-tests-and-arrives-on-vercel-s-ai-gat-ndoie9yjg) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"StepFun's Step 5 Preview: 600B MoE model scores 67/80 on coding tests and arrives on Vercel's AI Gateway","url":"https://daily.dev/posts/stepfun-s-step-5-preview-600b-moe-model-scores-67-80-on-coding-tests-and-arrives-on-vercel-s-ai-gat-ndoie9yjg","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/stepfun-s-step-5-preview-600b-moe-model-scores-67-80-on-coding-tests-and-arrives-on-vercel-s-ai-gat-ndoie9yjg"},"datePublished":"2026-10-08T17:19:28.236Z","dateModified":"2026-10-15T03:27:13.639Z","description":"StepFun released Step 5 Preview, a 600B-parameter mixture-of-experts model with 27B active parameters per token, supporting text and image input, a 1M-token...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/0f41fb2af592f695c6ccd6e30b39c2cd?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/0f41fb2af592f695c6ccd6e30b39c2cd?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":1,"discussionUrl":"https://daily.dev/posts/stepfun-s-step-5-preview-600b-moe-model-scores-67-80-on-coding-tests-and-arrives-on-vercel-s-ai-gat-ndoie9yjg","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":1}],"keywords":"ai,ai-coding,mixture-of-experts,ai-gateway","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"StepFun's Step 5 Preview: 600B MoE model scores 67/80 on coding tests and arrives on Vercel's AI Gateway"}]} {"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/stepfun-s-step-5-preview-600b-moe-model-scores-67-80-on-coding-tests-and-arrives-on-vercel-s-ai-gat-ndoie9yjg","comment":[{"@type":"Comment","text":"Does the 1M window change anything for screenshots, or does Step 5 resize images before they count? On Claude 4.7+ the long edge gets cut to 2576 px, so a 2540x18410 page capture reads as 355x2576 and 16 px text ends up about 2.2 px tall. A bigger context does not fix that. What is the image size cap here?","datePublished":"2026-10-09T07:31:03.532Z","url":"https://daily.dev/posts/NDOiE9yjg#c-tQG70gHfg","author":{"@type":"Person","name":"ab","url":"https://daily.dev/bickov","image":"https://lh3.googleusercontent.com/a/ACg8ocKwOmJ31UWqM4ZXB5ms-yRdFR4BYzyb7FPBRpg1O6DV0MIAF2EP=s96-c"}}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/stepfun-s-step-5-preview-600b-moe-model-scores-67-80-on-coding-tests-and-arrives-on-vercel-s-ai-gat-ndoie9yjg#faq","mainEntity":[{"@type":"Question","name":"What is StepFun's Step 5 Preview model and what are its specs?","acceptedAnswer":{"@type":"Answer","text":"Step 5 Preview is a 600-billion-parameter mixture-of-experts model from StepFun with 27 billion parameters active per token. It accepts text and image input, offers a 1-million-token context window, and can output up to 64K tokens. API pricing is $1 per million input tokens, $5 per million cached input tokens, and $2.70 per million output tokens. Full weights are planned for release on October 15th. daily.dev tracks new model releases like this for developers weighing coding-agent options."}},{"@type":"Question","name":"How did StepFun Step 5 Preview score on coding benchmarks compared to GLM 5.3?","acceptedAnswer":{"@type":"Answer","text":"Step 5 Preview scored 67 out of 80 (83.75%) across eight Kingbench 3 coding tasks run inside the Open Code agent, placing it above GLM 5.3 Flash and both MIMO v2.6 Pro and Flash scores, but behind GLM 5.3. Its strongest results were a fully playable archery game and a local fine-tuning workflow that trained and served a Gemma model, both scoring 9.5/10. Developers comparing coding models can follow benchmark results like these on daily.dev."}},{"@type":"Question","name":"How do I use StepFun's Step 5 Preview model in Claude Code or Cursor?","acceptedAnswer":{"@type":"Answer","text":"Step 5 Preview is available through Vercel's AI Gateway under the model name stepfun/step-5-preview and can be called from the AI SDK. It can also be configured in coding agents such as Claude Code, Codex, and Cursor through the Vercel CLI. AI Gateway adds no markup to inference pricing, including Bring Your Own Key requests, and offers usage reporting and API key budgets. daily.dev helps developers setting up new models in their coding agents stay current on options."}}]} ```