StepFun releases the open Step 3.5 Flash

Background

StepFun (阶跃星辰) is a Chinese AI company that builds large language models. Its Step series is the family to which this release belongs. Open-weights models, whose trained weights can be downloaded and run by others, are one way AI companies share their work, and this release is one such example.

What happened

In February 2026 StepFun released Step 3.5 Flash. It is a Mixture-of-Experts model with 196 billion total parameters and about 11 billion active parameters, and it is released under the Apache 2.0 license.^[1] The model is offered through StepFun's API platform and through OpenRouter, a service that gives access to many models through one interface.^[2] StepFun also said the model can run locally on high-end hardware for private deployment.^[4]

The technical report was posted to arXiv on 11 February 2026. It said the model's results were comparable to GPT-5.2 xHigh and Gemini 3.0 Pro.^[3]

Key details

The release supports tool use and a 256K-token context window.^[1][2] A context window is the amount of text, measured in tokens, that the model can take into account at once. A longer window lets the model work with larger documents or longer conversations.

The reported scores included 74.4 on SWE-bench Verified, a test of fixing real software bugs, and 97.3 on AIME 2025, a math competition benchmark.^[2] These figures come from StepFun's own reporting, so they describe the company's tests and not a neutral third-party review.

Impact

For readers who want to study or run a large model themselves, an open-weights release like this gives a concrete object to examine. The Apache 2.0 license means the weights can be used, modified and redistributed under permissive terms. The reported comparison with GPT-5.2 xHigh and Gemini 3.0 Pro shows the kind of claim AI companies now make about their open models, and readers should weigh it against independent evaluations rather than accept it as settled.