Step-3.5-Flash
Open-source MoE model for reasoning and agents
Step-3.5-Flash is an open-source foundation model from StepFun AI, designed for efficient reasoning and agentic tasks. It uses a sparse Mixture of Experts architecture to activate only 11 billion of its 196 billion parameters per token, enabling high speed and local deployment.
- Activates only 11B of 196B parameters per token
- Supports 256K context window with hybrid attention
- Scores 74.4% on SWE-bench Verified
- Generates up to 350 tokens per second in coding
- Runs on high-end consumer hardware locally
What it is
Step-3.5-Flash is a large language model developed by StepFun AI. Released in early 2026, it serves as a foundation model for reasoning and autonomous agent tasks. The system contains 196.81 billion parameters but activates only about 11 billion during generation. This design aims to balance the depth of proprietary systems with the efficiency required for real-time interaction. It supports a context window of 256,000 tokens and is optimised for coding and complex problem solving.
How it came about
The model was announced by StepFun AI in February 2026. It represents the company’s most capable open-source release to date. The development focused on creating a system that could rival closed-source competitors in reasoning depth while remaining accessible for local use. Technical details were published alongside the release, including performance benchmarks against other leading models.
Inside the model
The architecture uses a sparse Mixture of Experts approach. Each of the 45 transformer layers contains 288 routed experts and one shared expert. For every token, the model selects the top eight experts, keeping active parameters low. It employs Multi-Token Prediction to generate four tokens in a single pass, boosting speed. A hybrid attention mechanism combines sliding window and full attention layers to manage long contexts efficiently without excessive computational cost.
What it can do
The model achieves 74.4 per cent on SWE-bench Verified and 51.0 per cent on Terminal-Bench 2.0. It scores 97.3 on AIME 2025 and 98.4 on HMMT 2025 February. Generation throughput ranges from 100 to 300 tokens per second, peaking at 350 for coding. It handles multilingual tasks and supports agent loops with context management strategies for extended interactions.
How to use it
Weights are available openly via Hugging Face and ModelScope. The model can be run locally using backends like vLLM or llama.cpp on consumer hardware such as the Mac Studio M4 Max. API access is provided through StepFun’s platform and OpenRouter, which offers a free trial. Inference parameters suggest a temperature of 0.6 for chat and 1.0 for reasoning tasks.
Questions people ask about Step-3.5-Flash
How many parameters does Step-3.5-Flash have?
The model has a total of 196.81 billion parameters. However, it uses a sparse Mixture of Experts design, so only about 11 billion parameters are active for each token generated.
Can I run this model on my own computer?
Yes, it is optimized for local deployment. It runs on high-end consumer hardware like the Mac Studio M4 Max or NVIDIA DGX Spark using backends such as vLLM or llama.cpp.
What is the context length of the model?
Step-3.5-Flash supports a context window of 256,000 tokens. It uses a 3:1 Sliding Window Attention ratio to maintain performance across long documents while reducing computational overhead.
Good reading and sources
There is no English Wikipedia article on Step-3.5-Flash yet, so the facts on this page come from the sources listed above: huggingface.co/stepfun-ai/Step-3.5-Flash. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.