StepFun launches Step 5 Preview, a 600B MoE model with open weights slated for October

Announcement and target audience

On 20 September 2026 StepFun publicly introduced Step 5 Preview, branding it as a specialist large language model tailored for software development, engineering, professional workflows and financial analysis. The launch was simultaneous across the StepFun API and the newly released StepFun Studio web interface, both of which became reachable to external users on the same day.

StepFun positioned the model as a tool for developers who need deep code understanding, for analysts handling extensive financial reports, and for enterprises that require a high‑capacity system capable of processing very long documents without sacrificing response time.

Architecture and sparsity

The core of Step 5 Preview follows a sparsely‑gated mixture‑of‑experts (MoE) architecture. Although the total parameter count reaches 600 billion, the routing mechanism activates only about 27 billion parameters for each token, which dramatically reduces the compute load per inference while preserving the model’s overall expressive power.

This design choice mirrors earlier MoE systems that aim to scale capacity without a proportional increase in latency or energy consumption, and it enables Step 5 Preview to maintain competitive throughput even when handling the massive context windows advertised.

Multimodal inputs and context length

Step 5 Preview accepts both textual and visual inputs and produces text‑only outputs. The most striking technical claim is a context window of one million tokens, a magnitude that far surpasses the typical 8‑to‑32‑k token limits of most contemporary large language models. Such a window opens the door to ultra‑long‑form reasoning, extensive code‑base analysis, and the processing of full‑length financial statements in a single pass.

While the model can ingest images, it does not generate or edit visual content. The description of "text and images" suggests that vision‑language embeddings are merged into the same MoE routing network, allowing the model to reference visual information when producing its textual response.

Performance metrics and rankings

According to OrcaRouter’s Intelligence Index, Step 5 Preview achieved a score of 44 points, placing it 24th out of 200 evaluated models. The same ranking was reproduced in an OrcaRouter blog post dated 20 September 2026, confirming the model’s relative standing among both open and proprietary systems.

Leiphone reported that StepFun lists Step 5 Preview among the top three open models in the same index and claims that the per‑task cost is eight times lower than that of Claude Opus 5. The article notes that these cost comparisons are based on StepFun’s own pricing tables and have not been independently audited.

Open‑weight release timetable

StepFun announced that the model’s weights will become publicly available on 15 October 2026. Until that date, self‑hosting is impossible because the Hugging Face repository currently contains only a placeholder entry without any weight files, licensing details, model card, or configuration data, as confirmed by OrcaRouter’s verification routine.

Even after the weight files are uploaded, the lack of an accompanying license and configuration metadata means that users will need to wait for additional documentation before they can legally and technically integrate the model into their own pipelines.

• 600 billion total parameters
• 27 billion active per token
• 1 million token context window
• Supports text and image inputs

Artificial Analysis contributed the Intelligence Index score, while OrcaRouter supplied the ranking data. Both organisations operate independently of StepFun’s marketing materials and released their findings on the same day as the model announcement.

The claim of an eight‑fold cost reduction relative to Claude Opus 5 relies on StepFun’s internal pricing sheet. No third‑party audit of cost efficiency has been published, and the methodology differs from the performance‑based calculations used by OrcaRouter and Leiphone.

For English‑speaking organisations, the practical implication is that, starting mid‑October, they will gain access to a very large, sparsely‑gated model capable of handling extremely long contexts and multimodal inputs. If the advertised cost advantage proves accurate, it could meaningfully reduce expenses for code‑heavy workloads, large‑scale document analysis, and other professional tasks that benefit from extended context windows. Nonetheless, companies must still await the official license and configuration files before self‑hosting, and they should validate any cost claims against their own usage patterns.