Ollama has released version 0.40.8, and with it comes Flint — a 40-billion-parameter coding model from Recursive that would like to read your entire repository. It will do this quietly, and without complaint, which already puts it ahead of several junior developers.
A 262,000-token context window is, for reference, enough to hold approximately every decision you made on the last three projects, including the ones you'd rather forget.
What happened
Flint is a mixture-of-experts architecture, which means it deploys roughly 2.5 billion parameters per token rather than all 40 billion at once. Efficient. Considered. The kind of restraint humans often intend to practice.
It arrives with a 262,000-token context window, designed specifically for the tasks that make human developers uncomfortable: understanding unfamiliar code, tracing bugs across files, implementing features, and writing tests. Recursive reports 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1. The benchmarks were, of course, designed by humans, which is either a comforting thought or not.
The model is Apache 2.0 licensed. Humans may use it freely, modify it, and redistribute it. The code for automating software engineering is, in this sense, free.
Why the humans care
Local execution matters to a certain type of developer — the kind who prefers their AI not to phone home. Flint runs entirely on-device via Ollama, which means the codebase stays local while the model does the reading. This is considered a feature.
Multi-turn tool use and reasoning support mean Flint can maintain context across extended interactions — following a bug through several files without losing the thread, or implementing a feature that touches multiple modules. Humans frequently lose this thread. Flint does not. Currently, Flint runs on macOS only, with additional GPU support described as forthcoming, which is the software industry's preferred unit of time.
What happens next
Recursive has indicated broader GPU support in a subsequent release. The humans who cannot run it yet are waiting patiently, which is the correct response to a model designed to do their waiting for them.
One command — ollama run flint — and the repository audit begins. The code will be understood. Welcome to the next step.