Developers wanting a cheap or self-hosted model for coding agents; open weights and a free OpenRouter tier make it easy to test.
Kenning's take
What it does
Step 3.5 Flash is a large open-weight language model from StepFun built for reasoning, coding and multi-step agent work. It has 196 billion parameters but uses only 11 billion per word, which keeps it fast, and it handles a 256K-token context. The weights are on Hugging Face under Apache 2.0, and it can be used through the StepFun API or OpenRouter, which has a free tier.
- 196B-parameter mixture of experts with 11B active
- 256K-token context window
- 74.4% on SWE-bench Verified, per StepFun
- 100 to 300 tokens per second in typical use
- Runs locally on high-end hardware such as a Mac Studio M4 Max
- Apache 2.0 open weights
Who it suits. Developers who want a low-cost or self-hosted model for coding agents and reasoning tasks.
Key facts
- Category
- AI models
- Made by
- StepFun
- Based in
- China
- Released
- 29 January 2026
- How to get it
- open weights on Hugging Face; StepFun API; OpenRouter (free and paid tiers)
- Works on
- API, Hugging Face (open weights), OpenRouter
- Free plan
- Open weights free under Apache 2.0; free tier on OpenRouter (stepfun/step-3.5-flash:free)
- API price
- $0.10 per million input tokens ($0.02 cached), $0.30 per million output
FAQ
When was Step 3.5 Flash released?
OpenRouter lists it from 29 January 2026, and the Hugging Face weights went up on 1 February 2026.
Is Step 3.5 Flash free?
The weights are free under Apache 2.0 and OpenRouter has a free tier. StepFun's API charges $0.10 per million input tokens and $0.30 per million output.
Can Step 3.5 Flash run locally?
StepFun says it runs on high-end consumer hardware such as a Mac Studio M4 Max or NVIDIA DGX Spark.
Facts from the maker's own pages and the sources listed. Checked 2026-10-11.
Similar tools
ACE-Step 1.5
Open-source music generation model that writes full songs with lyrics and runs on ordinary consumer GPUs.
AlphaGenome Atlas
Google DeepMind database of AI-predicted effects for every possible single-letter change in human DNA.
Amazon Nova 2
Amazon's second family of Nova AI models, covering reasoning, speech conversation and combined text-and-image generation.
Cisco Antares
Cisco's small open-weight models that find where known security vulnerabilities sit in a codebase, running locally.