# Recursive Flint: A Free 40B Coding Model for Laptops, With 74.2% on SWE-bench Verified by Its Own Count

Richard Socher's Recursive released its first model under Apache 2.0. Only 2.5 billion parameters are active per token, so it runs at more than 125 tokens a second on an M4 Pro MacBook Pro. Here are the scores, the hardware it really needs, and the claims nobody has checked yet.

By Priya Nair (AI Industry Reporter) — published 2026-10-26, updated 2026-10-26
Source: https://aiscoutdaily.com/news/recursive-flint-free-local-coding-model-40b

## TL;DR
Recursive, the self-improving-AI startup led by former Salesforce chief scientist Richard Socher, released Flint this week, a free mixture-of-experts coding model with 40 billion total and 2.5 billion active parameters under the Apache 2.0 licence. Its Hugging Face card reports 74.2% on SWE-bench Verified, 89.7% on SWE-bench Pro v2 and 54.9% on Terminal-Bench 2.1, all run by Recursive. The company says it runs at more than 125 tokens a second on an M4 Pro MacBook Pro and over 150 on a 64GB Nvidia RTX Spark, and that its automated research system helped with nearly every step of building it. Recursive itself says Flint needs more guidance than large paid coding assistants.

Most local coding models are a compromise: small enough to run on a laptop, too weak to trust with a real bug. Recursive says Flint closes that gap. Coverage so far has focused on the claim that AI helped build it. For a developer deciding whether to install it, the scores, the memory it needs and the fine print matter more.

## The model in numbers

| Spec | Flint |
| --- | --- |
| Parameters | 40B total, 2.5B active per token (mixture of experts) |
| Licence | Apache 2.0, free to use and modify |
| SWE-bench Verified | 74.2% |
| SWE-bench Pro v2 | 89.7% |
| Terminal-Bench 2.1 | 54.9% |
| Speed, M4 Pro MacBook Pro | 125+ tokens per second (short contexts) |
| Speed, 64GB Nvidia RTX Spark | 150+ tokens per second (short contexts) |
| Context window | 262,144 tokens on the model card; the announcement says up to 512k, with 256k recommended |
| Full-precision download | About 80GB (BF16); Recursive's final release format is 4-bit NVFP4 |

*From Recursive's announcement and the Hugging Face model card. Benchmarks are Recursive's own, pass@1 averaged over at least five attempts.*

## What "runs on a laptop" really means

- Recursive tested on an M4 Pro MacBook Pro and a 64GB RTX Spark, and Wired's reviewer used a Mac with 48GB of memory. Neither the announcement nor the model card states a minimum memory figure.
- The 4-bit NVFP4 build is what makes laptop use possible. Recursive says the gap between it and full precision is within 0.7 points across its benchmarks.
- It works with llama.cpp, Ollama, Unsloth, LM Studio, vLLM and SGLang, and Recursive also released a small inference engine called rec.
- Running locally means it works offline, keeps code on your machine and has no usage fees.

## How it was built

Recursive says a small team built Flint from scratch in under four months, with its automated research system, the Eureka Machine, contributing to nearly every step from pretraining to inference. Socher told TechCrunch its contributions included the sparser model shape, low-bit training that shrank the model by 71% with almost no quality loss, and custom kernels. The company emerged from stealth in May with $650 million at a $4.65 billion valuation, backed by GV, Nvidia and AMD Ventures.

## What hasn't been checked

- Every benchmark and speed figure comes from Recursive. No independent leaderboard had published results at the time of writing.
- The context window is stated two ways: 262,144 tokens on the model card and up to 512k in the announcement.
- TechCrunch dates the release 21 October; the announcement page is dated 23 October.
- Recursive says Flint needs more guidance than large paid assistants and recommends breaking big tasks into smaller steps.

## What this means if you pick coding tools

- If you can't send code to a cloud model, Flint is worth testing first: Apache 2.0 and local inference remove the two usual blockers.
- Test it on your own repository before relying on its SWE-bench score. Vendor-run agent benchmarks vary a lot with the harness.
- Use it alongside a cloud assistant, not instead of one, for long multi-file changes, as Recursive's own guidance suggests.
- Check your machine's memory first. A 48GB or 64GB machine is the tested setup, not a typical laptop.

## Sources
- [Recursive: Introducing Recursive Flint (official announcement)](https://recursive.com/updates/introducing-recursive-flint)
- [Hugging Face: RecursiveSI/flint model card](https://huggingface.co/RecursiveSI/flint)
- [TechCrunch: Recursive launches Flint, a laptop coding model it says was largely built by AI (Oct 21, 2026)](https://techcrunch.com/2026/10/21/recursive-launches-flint-a-laptop-coding-model-it-says-was-largely-built-by-ai)