Step 5 Preview promises open weights for a 600B model

What this is about

Chinese AI company StepFun introduced Step 5 Preview on September 20, 2026. According to the company and multiple reports, it is a mixture-of-experts model with 600 billion total parameters, of which 27 billion are said to be active for each token. The API offers a context window of up to one million tokens, according to the announcement.

For developers, another promise matters just as much: StepFun says it will release the model weights on October 15, 2026. Until then, Step 5 Preview is mainly a hosted service. Pricing, benchmark results, and technical specifications largely come from the vendor or early measurements and should therefore be treated cautiously.

What Step 5 Preview actually does

A mixture-of-experts model does not activate the full network for every input. An internal routing mechanism sends individual tokens to a subset of specialized computational blocks. StepFun lists 600 billion total parameters and 27 billion active parameters per token. This design can reduce compute costs without limiting the model's entire knowledge base to the active subset.

The preview is available through an API. Reports list a price of $1 per million input tokens and $2.70 per million output tokens, plus a 95 percent discount for cached input. The long context window can hold large codebases, extensive document collections, or long agent sessions. A large context window does not guarantee that the model will use every passage reliably or remain consistent across very long inputs.

Why it matters

Competition is shifting away from model size alone toward three practical questions: how much performance developers receive per dollar, how expensive repeated input becomes, and whether teams can eventually operate the model themselves. StepFun addresses all three with low API pricing, a substantial cache discount, and a dated commitment to publish weights.

The weights have not yet been released. The word “open” also says nothing about the eventual license, training data, or full reproducibility. Businesses would need a permissive license; researchers would also benefit from the training recipe, evaluation data, and documented weaknesses. That promise cannot be fully assessed before October 15.

Early assessments place Step 5 Preview in the middle range of current frontier systems. Rankings are useful, but they do not replace an organization's own tests. Results may differ substantially for German-language work, long documents, tool use, and safety-critical workflows compared with an overall benchmark index.

In plain language

Think of Step 5 as a very large workshop. Instead of placing all 600 billion tools on the bench for every job, a foreman selects only the relevant drawers containing 27 billion tools. That saves effort, but the selection must be correct: if the foreman opens the wrong drawer, the size of the workshop does not help.

A practical example

A software team wants to analyze a 700,000-token codebase and then inspect the same files with small changes across 20 iterations. The one-million-token window could hold the initial codebase in one request. If the provider applies the stated 95 percent discount to unchanged cached input, repeated context could cost substantially less than the first pass.

This is not a cost forecast. Tokenization, cache rules, output length, and reprocessing can change the bill. The team would need a representative test with real repositories, defined quality thresholds, and a decision on whether confidential code may leave its own environment at all.

Scope and limits

First, the open weights are only promised. Until release, the license, download size, hardware requirements, and local runtime remain unknown. Second, most performance claims currently come from vendor information and early aggregators; independent testing for German, long context, and tool use is still limited. Third, a capacity of one million tokens does not mean one million tokens of reliable attention. Important details may be missed in long inputs.

There are operational limits as well. A model with 600 billion total parameters is likely to require substantial memory and infrastructure despite sparse activation. Small teams should not assume that “open weights” means “runs on a laptop.” For sensitive data, users must also review API location, contractual terms, logging, and deletion periods.

SEO & GEO keywords

StepFun, Step 5 Preview, mixture of experts, 600B model, open model weights, open-weight AI, one-million-token context, AI API, model pricing, local AI deployment, Chinese AI models