Step 5 Preview
Open-weights model for long-horizon coding, frontend, and finance work
About model
Step 5 Preview is StepFun's open-weights model for long-horizon agentic tasks, performing especially well in AI coding, software engineering, professional knowledge work, and finance. A sparse Mixture-of-Experts model with 600B total parameters and 27B active per token, it uses a 92-layer narrow-and-deep architecture that gives the model a deeper compute path for multi-hop reasoning while keeping parameter count and compute cost in check. Sparse GQA with block-level indexing keeps a 1M-token context window affordable, and the model accepts native text and vision input. Reinforcement learning across million-scale verifiable environments, with context compression built into policy training, teaches the agent to organize task state and keep going over long trajectories. Available on Together AI.
600B
Sparse MoE built for long-horizon agentic work
1M
Large codebases and long documents in a single prompt
Native
Screenshots, mockups, and documents alongside text
- Real Engineering Work: Fixes bugs in unfamiliar codebases, implements features across files, refactors and debugs, with changes that run and pass checks
- Design to Product: Turns screenshots, mockups, or PRDs into working web pages, or builds playable browser games from assets
- Professional Deliverables: Analyzes revenue, profit, cash flow, and risk from annual reports and industry material, then delivers reports, editable spreadsheets, and slide decks
- Production-Ready Infrastructure: 99.9% SLA, available on Together AI dedicated infrastructure
Model card
Architecture Overview:
• Sparse Mixture-of-Experts with 600B total parameters and 27B active per token
• 92-layer narrow-and-deep architecture providing a deeper compute path for implicit multi-hop reasoning while keeping parameter count and compute cost in check
• Sparse GQA combines block-level indexing with sparse computation to cut the cost of million-token contexts; 8-token block indexing reduces indexing and top-k selection overhead to one-eighth
• 1M-token context window with native text and vision input
Training Methodology:
• FP32 precision retained for key states and parameters, with fine-grained gradient clipping, to curb activation blow-ups and gradient spikes in a deep network
• Million-scale verifiable environments spanning software engineering, scientific reasoning, ML R&D, and professional knowledge work, with task difficulty, diversity, and evaluation reliability continuously calibrated
• Reinforcement learning for long-horizon tasks with context compression built into end-to-end policy training, plus finer-grained reward attribution across long trajectories
Performance Characteristics:
• Designed for long-horizon agentic tasks, with particular strength in AI coding, software engineering, professional knowledge work, and finance
• Compatible with Step Code, StepFun's open-source agent harness built around the model
Prompting
Together AI API Access:
• Access Step 5 Preview via Together AI APIs using the endpoint stepfun-ai/Step-5-Preview
• Authenticate using your Together AI API key in request headers
• Supports text and image input with text output across a 1M-token context window
• Available on Together AI dedicated infrastructure
Applications & use cases
Agentic Software Engineering:
• Fix real bugs in unfamiliar repositories end to end, with changes that run and pass checks
• Implement features across files and refactor large codebases held in a single 1M-token context
• Run long-horizon agent sessions that keep task state organized across extended trajectories
Frontend & Interactive Builds:
• Turn screenshots, mockups, or product requirement documents into working web pages
• Build playable browser games from supplied assets
• Iterate on interfaces with visual input grounding each revision
Finance & Professional Knowledge Work:
• Analyze annual reports and industry material for revenue, profit, cash flow, and risk
• Deliver financial analysis as reports, editable spreadsheets, and slide decks
• Process long filings and document sets in one prompt
- Model providerStepfun
- TypeChatReasoningVision
- DeploymentDedicated
- Parameters600B
- Activated parameters27B
- Context length1M
- Input modalitiesTextImage
- Output modalitiesText
- ReleasedOctober 14, 2026
- External link
- CategoryChat
