Recursive Flint: The Free Coding Model That Runs on a Laptop

Recursive Flint is a coding model you can download for free and run on your own laptop. Recursive, the AI lab led by Richard Socher, released it on 21 October 2026 under the Apache 2.0 licence, with the weights on Hugging Face. The headline claim is unusual: the company says Flint was largely built with the help of its own AI research system. That claim is worth reading carefully, and so is the more useful fact underneath it. A capable coding model now fits on a machine many developers already own.

This post covers what Flint is, what you need to run it, what the benchmarks do and do not tell you, and the parts of the launch that deserve some caution.

What Flint is

Flint is a mixture-of-experts model. The model card gives 39.8 billion parameters in total, of which about 2.5 billion are active for any one token. That split is why it is fast on modest hardware: the model is large, but each step only uses a small slice of it. It takes text in and gives text out, supports step-by-step reasoning, and can call tools, which is what a coding agent needs to read a codebase, run tests and edit files.

Recursive describes the intended job plainly: investigating codebases, fixing bugs, implementing features and writing tests. It is not a chat assistant for general questions, and it is not multimodal. If you want a sense of how a model differs from the agent wrapped around it, AI Agent vs LLM sets out the distinction.

What you need to run it

The full-precision checkpoint is about 80 GB, and the model card points it at a single data-centre GPU. That is not the laptop version. For local use Recursive publishes quantised variants, and its announcement lists support in llama.cpp, Ollama, Unsloth, LM Studio, vLLM and SGLang, plus a small engine of its own called rec for recent Apple Silicon.

The speed figures come from Recursive’s own tests: more than 125 tokens per second on an M4 Pro MacBook Pro, and more than 150 on a 64 GB NVIDIA RTX Spark at short contexts. TechCrunch reports that a Wired reviewer ran it on a Mac with 48 GB of memory. Treat a machine of that class as the realistic starting point, not a base-model laptop.

One number is inconsistent between the two official pages. The announcement says Flint supports up to 512k tokens of context and recommends staying within 256k. The model card documents a configuration of 262,144 tokens. If you plan to feed it an entire repository, assume the smaller figure until Recursive reconciles them.

The benchmarks, read with care

The model card reports 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified. Recursive says Flint leads mixture-of-experts models under 50 billion parameters on all three. These are Recursive’s own release evaluations, run on its own setup, and the card itself warns that evaluation setups vary and that some comparison scores are developer-reported.

That does not make them wrong. It makes them a reason to try the model on your own work rather than a verdict. A benchmark tells you how a model did on someone else’s tasks. Your codebase, your conventions and your test suite are the only test that matters, and the habit of checking output against something you trust is covered in How to Check an AI Answer When You Are Not the Expert.

What “largely built by AI” means

Recursive says its automated research system, which it calls the Eureka Machine, helped with nearly every stage of Flint’s development. According to TechCrunch, the system proposed a sparser model shape, low-bit training that shrank the model by 71% with little quality loss, and custom kernels, and it let a small team build Flint in under four months.

Two things are missing from that account. Recursive has not said how much researcher time the system actually saved, and the system is supervised by people throughout. It is an AI research assistant that runs experiments and suggests what to try next, not a model that improved itself unattended. That still matters, because it is a public, downloadable example of AI doing a real share of AI research. It is just a smaller claim than the headline suggests. For context on how models are built in the first place, see How to Create an AI Model.

Why a local model is the interesting part

For most people the practical change is not the AI-built story. It is that the model runs on your machine. Code you send to a hosted assistant leaves your laptop. Code you give a local model does not, which is why TechCrunch notes Flint can work offline and keep code private. For anyone whose employer forbids pasting source code into a cloud chatbot, that is the difference between using an AI coding tool and not using one.

One caveat from the model card is easy to miss: hosted API terms and data practices are separate from the weight licence. If you use Flint through someone’s cloud service rather than on your own hardware, the privacy argument depends on that provider’s terms, not on Apache 2.0. The Apache 2.0 licence itself is permissive: commercial use and modification are allowed.

The limits Recursive admits

• —It needs more guidance than the large paid coding assistants. Recursive’s own advice is to break big tasks into smaller steps with explicit instructions.
• —It can write faulty code and make inappropriate tool calls. The model card says to test what it produces and review consequential changes.
• —An agent that can run commands needs its own permissions and boundaries. The model does not supply them.
• —Long reasoning adds latency and can use up the output budget before an answer is finished.

Those are honest limits, and they are the same ones that apply to every coding model, just more visibly. A tool that can act on your machine deserves the same care as a new colleague with admin rights. What Happens When an AI Agent Takes Control of Your Computer walks through what that means in practice, and OpenAI’s recent report on agents weakening their own monitoring is a reminder of why boundaries matter.

Should you try it?

1. 01If you write code and have a recent Mac or a well-specified PC, yes. It costs nothing but disk space and an afternoon.
2. 02Start with a task you already know the answer to: a small bug, a test you would have written anyway. Compare its work to yours.
3. 03Keep tasks small. Recursive says so itself, and it is good practice with any coding model.
4. 04If you cannot run code locally, or your work does not involve code, Flint is not for you yet. A hosted assistant will serve you better.

If you are weighing it against a paid tool, Muse Code vs Claude Code shows what the hosted options offer, and Agentic Workflows explains how to break work into the small, checkable steps that a model like Flint handles best.

Whatever model you end up using, the skill that carries over is the same: give clear instructions, keep the task small, and check the result before you trust it. That is what Coursium teaches, on your phone, for people who use AI at work.