Step 3.5 Flash

What it is

Step 3.5 Flash is a large language model from StepFun (阶跃星辰). It is an open sparse Mixture-of-Experts (MoE) model with 196 billion total parameters, of which about 11 billion are active for each token processed.^[1] A MoE model is split into many expert sub-networks, and a routing step sends each token to only a few of them. "Sparse" means that only part of the model does the work for any one token, which keeps the compute per token much lower than the total size would suggest.

In detail, the model has 45 layers. Each layer contains 288 routed experts and one shared expert that every token uses, and the router sends each token to the top 8 routed experts.^[1]

Origins

Step 3.5 Flash belongs to StepFun's Step series of models. StepFun released it in February 2026, and the technical report was posted to arXiv on 11 February 2026.^[3] StepFun called it its most capable open-source foundation model, aimed at reasoning and agentic work at lower compute cost.^[2] An agentic model is one designed to carry out multi-step tasks, such as using tools and working through a sequence of actions, rather than only answering a single question.

Main features

The model uses a hybrid attention design. For every full-attention layer, which looks at all previous tokens, there are three sliding-window layers, which look only at a nearby window of tokens. This reduces the cost of long inputs.^[1] It also uses 3-way multi-token prediction (MTP-3), which lets the model predict several tokens at once during generation to decode faster.^[1]

The context window is 256K tokens, and the model is released under the Apache 2.0 license.^[1] It supports tool use.^[1][2] StepFun reports that typical generation runs at 100 to 300 tokens per second, peaking at about 350 tokens per second for single-stream coding. The company says the model can run locally on hardware such as a Mac Studio with an M4 Max chip or an NVIDIA DGX Spark.^[2]

StepFun also acknowledged that Step 3.5 Flash uses longer generation trajectories than Gemini 3.0 Pro to reach similar quality.^[2]

Reported results

StepFun published scores on several benchmarks. These are the company's own figures:

• SWE-bench Verified, a test of fixing real software bugs: 74.4.^[1]
• Terminal-Bench 2.0, a test of working in a command-line terminal: 51.0.^[1]
• tau2-Bench: 88.2.^[1]
• AIME 2025, a math competition benchmark: 97.3.^[1]
• LiveCodeBench-V6, a coding benchmark: 86.4.^[1]
• GPQA Diamond, a set of difficult science questions: 83.5. MMLU-Pro, a harder version of a broad knowledge test: 84.4.^[1]

StepFun also reported an average of 81.0 across eight benchmarks. For comparison, its figures give 82.2 for GPT-5.2 xhigh, 80.7 for Gemini 3.0 Pro, 80.6 for Claude Opus 4.5 and 77.3 for DeepSeek V3.2.^[2] Averages over different benchmark sets are sensitive to which tests are included, so these figures are best read as the company's own comparison.

Where it stands

Step 3.5 Flash is an example of an open-weights model that targets reasoning and agent tasks while keeping the active parameter count and per-token compute modest. The model is offered through StepFun's API platform as step-3.5-flash and through OpenRouter.^[1] Anyone who wants to check its claims can compare the published model card and technical report with independent evaluations, and can run the open weights if their hardware allows.