# Recursive launches Flint, a 40B open-weight coding model

> Flint offers 40 billion parameters, 2.5 billion active, and runs at 125 tokens/sec on a MacBook Pro.

Oossa · 2026-10-21 · https://oossa.com/en/recursive-launches-flint-a-40b-open-weight-coding-model

Recursive has made its new coding model, Flint, available under an Apache 2.0 license. Flint is a mixture‑of‑experts (MoE) model with a total of 40 billion parameters, but only 2.5 billion are active for any given input. The company says the model reaches 74.2 % accuracy on the SWE‑bench Verified benchmark, a test of code‑writing ability. Recursive also reports that Flint can generate more than 125 tokens per second on an M4 Pro MacBook Pro.

## What does Flint mean for developers?

Because Flint is open‑weight, anyone can download the model weights and run it locally, without needing a cloud service. That can lower costs for developers who need a fast coding assistant on personal hardware. The reported speed of 125 tokens/sec means a typical code snippet can be produced in a few seconds on a recent MacBook.

## The facts

- Flint has 40 billion total parameters and 2.5 billion active parameters.
- The model is released under the Apache 2.0 license.
- Recursive reports 74.2 % accuracy on the SWE‑bench Verified benchmark.
- Flint runs at over 125 tokens per second on an M4 Pro MacBook Pro.

## Why it matters

Developers can now run a high‑capacity coding model on their own laptops, avoiding subscription fees for cloud AI services. The open‑weight release also lets researchers experiment with the model without legal restrictions. Performance on a consumer laptop suggests Flint is practical for everyday coding help, though independent speed and accuracy tests are not yet available.

## Sources & references

1. [Recursive Releases Flint, an Open-Weight 40B MoE Local Coding Model for Agents with 2.5B Active Params](https://www.marktechpost.com/2026/10/21/recursive-flint-local-coding-model/) – MarkTechPost, 2026-10-21

Last updated: 2026-10-21