Recursive Flint: a free coding model that runs on a laptop
Recursive Flint is an open MoE coding model (40B, 2.5B active) that runs even on a laptop. Most of the work on it was done by AI.
This article in other languages:العربيةDeutschEspañolFrançaisItaliano日本語Polski简体中文
In this article
On October 21, the startup Recursive released Flint, an open coding model under the Apache 2.0 license that runs on an ordinary laptop. According to TechCrunch, the company says an automated research system, the so-called “Eureka Machine”, helped with almost every stage of building Recursive Flint.
What is Recursive Flint?
It’s a MoE (mixture-of-experts) model with 39.8 billion parameters, of which about 2.5 billion work on each token. Hence the “40B-A2.5B” label on the model card. Under the hood, among other things:
- 49 layers, 47 of which are MoE layers with 256 experts, 6 of which are picked for each token, plus one shared expert that is always active.
- A context window of 262,144 tokens (prompt and output combined).
- A knowledge cutoff of June 2026.
- The Apache 2.0 license, so we can use it commercially too.
The full BF16 checkpoint weighs about 80 GB. We won’t run it on a laptop in that form, but according to the model card, Recursive prepared separate quantized versions for NVIDIA chips and Apple Silicon. Hugging Face already has more than a dozen of them, with ready-made links to llama.cpp, LM Studio, Jan and Ollama.
What can Flint do?
Results from the model card (reported by Recursive):
| Benchmark | Score |
|---|---|
| SWE-bench Pro v2 | 89.7% |
| SWE-bench Verified | 74.2% |
| Terminal-Bench 2.1 | 54.9% |
Terminal-Bench was run with the Terminus 2 harness and SWE-bench with mini-swe-agent, averaging at least five attempts. These are the vendor’s own numbers, so let’s take them with a grain of salt until independent measurements show up.
Speed is the more interesting part. According to Recursive’s tests, on a MacBook Pro with an M4 Pro, Flint generates about 131 tokens per second, and over 150 on an NVIDIA RTX Spark. There’s also a first independent opinion: according to TechCrunch, Will Knight of Wired built small programs with Flint on a Mac with 48 GB of memory, and the model “nailed his coding requests every time”.
A model built by AI?
According to TechCrunch, a small team built it from scratch in under four months, and the Eureka Machine reportedly proposed, among other things, a sparser model shape, custom kernels and low-precision training, which shrank the model by 71% with almost the same quality.
Recursive hasn’t said how much researcher time this saved, and according to the article, the current version of Eureka is a research assistant supervised by humans, not an autonomous system. The long-term goal of CEO Richard Socher (formerly chief scientist at Salesforce) is full recursive self-improvement of AI, but with humans who can stop the process.
The company came out of stealth in May with $650 million in funding at a $4.65 billion valuation, and its investors include GV, NVIDIA and AMD Ventures.
How to try Recursive Flint?
The easiest way is one of the quantized versions on the model page, in LM Studio or Ollama. I described how to do it step by step in my guide on how to run AI locally. The recommended settings from the model card are temperature 1.0, top-p 0.95, top-k 20, and at least 32,768 tokens per response for hard tasks.
For the full version, the model card describes a single B300 or H200 GPU with SGLang and an OpenAI-compatible API. We can plug an endpoint like that into an agent harness, just like free DeepSeek in OpenCode back in the day.
One caveat from Recursive itself: Flint needs more guidance than the big paid coding assistants, so it’s better to split large tasks into smaller steps.
FAQ
Is Recursive Flint free?
Yes. The weights can be downloaded from Hugging Face under the Apache 2.0 license, and running it locally only costs us electricity (and hardware wear).
How much memory does Flint need?
The full BF16 version weighs about 80 GB and, according to the model card, is meant for a B300 or H200 GPU. On a laptop we use the quantized versions. Wired tested the model on a Mac with 48 GB of memory.
Was Flint really built by AI?
According to Recursive, the Eureka Machine helped with almost every stage, but under human supervision. The company hasn’t said what share of the work the system did on its own.