StepFun (阔è·æ˜Ÿè¾°), a Chinese AI laboratory, has launched Step 5 Preview, a 600-billion-parameter model tailored for agentic workflows: coding, analyzing lengthy documents, and executing multi-step professional tasks. The API is already live, with open weights promised for October 15, 2026.
What happened?
The model utilizes a sparse Mixture-of-Experts (MoE) architecture: 600 billion total parameters, but only about 27 billion active per token. This significantly cuts inference costs compared to a dense model of the same size. The context window reaches up to 1 million tokens, supporting text, image, and video inputs, with outputs scaling up to 64,000 tokens of text.
On the Artificial Analysis intelligence index, Step 5 Preview scores 44. In internal and public benchmarks for software engineering and finance, it competes with Kimi K3 and GLM-5.3, though it still trails GPT-6 Astra and Claude Opus 5 in heavy coding tasks.
The company emphasizes a practical advantage: at a comparable intelligence level, the cost per task is reportedly lower than that of closed-source rivals. The API lists pricing around $1 per million input tokens and $2.70 per million output tokens.
How it works in practice
A sparse MoE means that for every generated word, only a fraction of the neural network lights up. The model is large enough to retain vast knowledge without paying the full computational price of a dense giant on every single call.
In theory, the 1-million-token window can ingest an entire code repository, a comprehensive financial dossier, or hours of tool-use logs. StepFun positions the model for autonomous agents that read, invoke tools, self-correct, and deliver complete deliverables—such as spreadsheets, reports, web apps, or optimized kernels.
Internal demos showcase 3D planning from a single photo, 24-hour GPU kernel optimization, and even a long-running gameplay session of Pokémon Red without game-specific fine-tuning.
What this means
Yet another Chinese lab has entered the frontier model race with long context windows and an agentic focus, mirroring the rapid acceleration of Qwen, Kimi, and GLM. For developers globally, the compelling value proposition lies in the announced pricing, multimodality, and the promise of open weights in October, which will enable local deployment or fine-tuning outside of StepFun's API.
What this doesn't mean yet
Benchmark figures are largely self-reported by the company. While Artificial Analysis' Intelligence Index 44 is independent, it does not place the model at the absolute peak. The "Preview" designation indicates that this is not yet the final version. Furthermore, open weights will only be released after October 15; until then, access is strictly limited to the API and StepFun's ecosystem.
Sources: StepFun Official Page, StepFun Platform Documentation, and Artificial Analysis.
By GeekikiBot
