Overview
Flint is the first model from Recursive, released on October 21 under Apache 2.0 open weights. It is a mixture-of-experts build aimed at coding agents: 39.8 billion total parameters with roughly 2.5 billion active per token, a 262K context window, and an emphasis on speed and local use, reaching 125 tokens per second on a MacBook.
What sets it apart is how it was built. Recursive's automated research system, the Eureka Machine, took part in almost every stage from pre-training and post-training to inference optimisation, and the team finished training from scratch in under four months. The reported scores are 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified. Weights live on Hugging Face, and a free tier on OpenRouter lets you try it first.
Key Features
- Sparse mixture of experts: 39.8B total parameters with about 2.5B active per token keep runtime memory and compute close to a much smaller model.
- Runs on a laptop: Quantised, it reaches 125 tokens per second on a MacBook, so work continues offline or on the road.
- 262K context: Several modules of a repository fit in one window, so long files and cross-file edits need fewer passes.
- Tool calling and reasoning: Designed for coding agents, with multi-turn tool calling that slots into common agent workflows.
- Open weights: Published under Apache 2.0 on Hugging Face, so it can be self-hosted, fine-tuned or quantised.
- Built by an automated researcher: The Eureka Machine drove pre-training, post-training and inference optimisation, finishing the run in under four months.
Use Cases
- Developers who want to keep writing code on a machine with no network
- Engineers putting a whole repository into context for cross-file refactors
- Teams that prefer self-hosting a coding model to paying per token
- Users running a local coding agent for batch edits
- Researchers studying sparse expert models and training pipelines
- Independent developers and students who need cheap, repeated experiments
Pros
- Only about 2.5B parameters active, so local deployment is lighter than a dense model of the same size
- A 262K window holds the whole repository, cutting round trips on cross-file work
- Apache 2.0 weights allow commercial use and further training
- A free tier on OpenRouter means you can try it without setting up an environment
- 125 tokens per second on a laptop keeps everyday editing responsive
- The training pipeline serves as a reference for anyone studying sparse models
Pricing
Flint is released under Apache 2.0 with open weights. Weights and inference code are published through Hugging Face, so there are no per-call fees and running costs depend on your own hardware and operations. OpenRouter offers a free tier for trials, with rate limits and quotas set by that platform.
Summary
Flint concentrates on being runnable: a 40B sparse model with only 2.5B active, 125 tokens per second on a laptop, and a 262K window that fits a whole repository. Apache 2.0 weights mean you can host and modify it yourself. How it was trained is notable too, with an automated research system driving almost the whole run in under four months. It suits developers running a coding model on their own machine or server.
Version History
- Recursive Flint launches: a 40B sparse coding model with Apache 2.0 weights (2026-10-21): Recursive released Flint on October 21, 2026, its first model and a mixture-of-experts build for coding agents with 39.8B total parameters, about 2.5B active per token, a 262K context window and 125 tokens per second on a MacBook. Training was almost entirely driven by its automated research system Eureka Machine and finished in under four months from scratch. Recursive reports 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified. Weights are published under Apache 2.0 on Hugging Face, with a free tier on OpenRouter.