Editorial illustration for StepFun's Step 5 Preview Offers 1M-Context, 600B-Parameter MoE Model
StepFun's Step 5 Model Hits 600B Parameters, 1M Context
StepFun put a number on the table Tuesday: 600 billion parameters, 27 billion of them active per token, wrapped in a Mixture-of-Experts model called Step 5 Preview. The Chinese AI lab is aiming it at software engineering, finance, and other professional knowledge work where long-horizon agentic tasks matter more than chat-style back and forth. The pitch isn't raw scale for its own sake. StepFun frames Step 5 as sitting on a "Pareto frontier," meaning it claims comparable intelligence to bigger, pricier models at a fraction of the per-task cost.
Right now that's only testable through StepFun's hosted API and platform. Self-hosting isn't on the table yet. Open weights are promised for October 15, 2026, and when they land, the hardware math gets serious fast: 600B parameters in BF16 alone runs close to 1.2 terabytes before you even account for KV cache, which means multi-GPU server setups will be mandatory for anyone running it locally.
The model handles a 1M-token context window, takes text, images, and video as input, and outputs text with adjustable reasoning effort. StepFun also built in tool calling, JSON Schema support, and a Claude Code integration through its Step Plan.
Step 5 Preview is a sparse Mixture-of-Experts (MoE) model. It holds about 600B total parameters and activates about 27B per token. That is roughly 4.5% of the weights per token.
Why this matters
For teams building agentic tools, Step 5 Preview is worth watching mainly for the cost claim, not the parameter count. A 600B-total, 27B-active MoE with a 1M-token window and multimodal input is a reasonable spec sheet, but StepFun's whole pitch rests on "comparable intelligence, lower cost," and that's a number we haven't seen independently checked yet. Anyone weighing it against Gemini or Qwen for long-context coding work should ask for real task-level pricing, not marketing math.
The October 15 open-weights date is the detail we'd flag hardest. Until then, this is API-only, which means StepFun controls the story on latency, uptime, and actual per-task spend. That's fine for a quick pilot, but not something founders should build production infra around without a fallback plan. Once weights ship, self-hosting will let researchers actually test the "Pareto frontier" claim against real workloads instead of StepFun's own benchmarks.
Reasoning-effort tiers and JSON Schema support suggest StepFun is chasing the same developer-experience checklist as OpenAI and Anthropic. Whether it's a genuine cost alternative or just another entrant in an already crowded field depends entirely on what happens after October 15.
Related Reading
Business & Startups Nordic pilot adds Gemini for Education, NotebookLM to boost AI literacy
Business & Startups Kling launches Video O1, all-in-one model with MVL bridge using transformer
Business & Startups DeepSeek Seeks More Capital Weeks After USD 7B Funding Round
From Archives Anthropic Delays IPO After OpenAI Despite Growing Enterprise Revenue
From Archives Unity Launches Official Plugins for Claude Code and OpenAI Codex
Common Questions Answered
What are the key specifications of StepFun's Step 5 Preview model?
Step 5 Preview is a sparse Mixture-of-Experts model with 600 billion total parameters and 27 billion active parameters per token, representing approximately 4.5% of weights activated per token. The model features a 1 million token context window and is designed for professional knowledge work including software engineering and finance applications.
How does Step 5 Preview's MoE architecture compare to larger models in terms of efficiency?
StepFun positions Step 5 Preview on a Pareto frontier, claiming it delivers comparable intelligence to larger models while maintaining greater efficiency through its sparse Mixture-of-Experts design. By only activating 27 billion of its 600 billion parameters per token, the model achieves cost advantages without sacrificing performance on long-horizon agentic tasks.
What are the primary use cases StepFun is targeting with Step 5 Preview?
StepFun is aiming Step 5 Preview at software engineering, finance, and other professional knowledge work where long-horizon agentic tasks are more important than chat-style interactions. The model's 1 million token context window makes it particularly suitable for complex, extended reasoning tasks in these domains.
Why is cost-effectiveness more important than raw parameter count for Step 5 Preview?
For teams building agentic tools, Step 5 Preview's main value proposition is its cost efficiency rather than its total parameter count. StepFun's pitch centers on delivering comparable intelligence to competitors like Gemini and Qwen at a lower cost, making pricing verification crucial for potential users evaluating the model.
Further Reading
- StepFun Launches Step 5 Preview: 600B Sparse MoE, 1M Context, Weights Open Oct 15 - Pandaily
- Step 5 Preview: Advancing the Pareto Frontier - 阶跃星辰 - StepFun
- Step 5 Preview: StepFun's 600B Agentic Model - AI/TLDR
- Step 5 Preview — StepFun's 600B flagship for agent work - AI/TLDR
- Step 5 Preview API, Pricing & Playground - Vercel AI Gateway
