Ollama v0.40.8 runs Flint, Recursive's open coding model for agents, on Macs first
$ ollama run flint
Ollama now runs Flint, the coding model from Recursive. It is built for the work coding agents do inside a repository: reading unfamiliar code, tracing bugs, adding features and writing tests.
Flint runs on macOS now, and other GPUs get it in a later release
Flint runs on Ollama's MLX engine, so this release supports it on macOS. Support for other GPUs comes in a later release. If your Ollama machine runs on another GPU, Flint has to wait.
Flint has 40B parameters and puts about 2.5B to work on each token
Flint is a mixture of experts with 40 billion parameters, and only about two and a half billion of them work on each token. The default download is 23 gigabytes, under the Apache 2.0 license.
Flint holds 262K tokens for tasks that run across many files
- Reasoning before it answers
- Multi-turn tool use inside an agent loop
- 262K-token context window
Flint reasons before it answers, calls tools over many turns, and holds a context window of about 260 thousand tokens. That leaves room for a long task that moves across files and tool calls.
Flint fixes about three in four real GitHub issues on SWE-bench Verified
Given real bug reports from open source Python projects, Flint writes a working fix for about three in four. At a command line, it finishes just over half of the jobs it gets.
Which coding agent would you run Flint in?
On a Mac, Flint is one pull away. Which coding agent would you run it in?
