Intel / post

Step 3.5 Flash: 196B MoE open weights built for agent workloads

Source: x.com
Date: Published Feb 2, 2026

Context

StepFun released Step 3.5 Flash as an foundation model built for agent and coding workloads. It uses a sparse (MoE) architecture with 196 billion total parameters but only 11 billion active per token, meaning most of the model's parameters sit unused for any given request, which is what lets a model this large run at the 100 to 300 tokens per second StepFun reports, with a 350 tokens per second peak on single-stream coding tasks.

StepFun reports 74.4% on , a benchmark of real GitHub issues that measures whether a model can produce a working code fix, and 51.0% on Terminal-Bench 2.0, which tests command-line agent tasks. Both figures are StepFun's own reported evaluation results rather than scores independently reproduced by a third party. The model also supports a 256K-token , and StepFun says it runs locally on hardware such as a Mac Studio M4 Max or an Nvidia DGX Spark, positioning it for on-device rather than only cloud deployment.

Terms in this piece · Glossary

open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
SWE-bench — The standard benchmark for AI coding agents: real GitHub issues from real repositories, scored by whether the agent's patch passes the project's own tests.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.

Read the source x.com