How to run Flint, the free coding model that runs on a laptop

What Flint is

Flint is Recursive’s first model, released on October 21, 2026. Its model card describes a mixture-of-experts model with 39.8 billion parameters in total and about 2.5 billion active for each token: 256 experts, six chosen per token, plus one that always runs. That is how a model this capable fits on a laptop. Most of it sits in memory while only a small part does the work for each word.

The weights are on Hugging Face under Apache 2.0, a permissive license that allows commercial use. Recursive says its automated research system helped with nearly every step of building it, which is most of what the launch coverage was about.

How good it is

Recursive’s own numbers, from the model card:

Benchmark	Flint
SWE-bench Verified	74.2%
Terminal-Bench 2.1	54.9%
SWE-bench Pro v2	89.7%

Each is an average over at least five attempts. These are the publisher’s results; we haven’t found independent evaluations yet.

Recursive says Flint leads mixture-of-experts models under 50 billion parameters and rivals dense models of 25 to 40 billion while running more than five times faster. It isn’t the best small model on every count: TechCrunch reports that Alibaba’s Qwen3.8 27B scored higher on Recursive’s six-test average. Where Flint stands out is speed: more than 125 tokens a second on an M4 Pro, by Recursive’s measure over a 128k-token session.

Recursive is also plain about its limits. It “may need more guidance” than the large paid assistants, and works best when you break big tasks into smaller steps.

What you need to run it

• On a Mac with Recursive’s runtime: an M4 Pro, M5 Pro or M5 Max with at least 48 GB of unified memory and about 30 GB of free disk, according to Recursive’s docs.
• With Ollama: the default flint:40b build is a 23 GB download. Recursive doesn’t publish a memory figure for it, so leave room beyond the file size for the context.
• With llama.cpp or LM Studio: community quantizations range from about 11 GB at the smallest to 42 GB at 8 bits. Smaller builds trade away quality.
• On a GPU server: vLLM and SGLang serve the full model, which is about 80 GB.

Three ways to run it

1. Ollama. Install Ollama, then run ollama run flint. Recursive’s docs name a specific build, flint:40b-a2b-qat-int4, the same 23 GB. Ollama’s page tags the model for tools and thinking, which an agent needs.
2. LM Studio. Recursive’s guide uses lms get recursivesi/flint, then loads it and starts LM Studio’s local server on port 1234, an OpenAI-compatible endpoint other apps can call.
3. Recursive’s rec, on a Mac. rec is Recursive’s own runtime. It serves Flint at http://127.0.0.1:11500/v1, and rec launch opencode starts OpenCode already pointed at it.

Give it real work

A chat window is the least of what Flint can do. Its card says it handles OpenAI-compatible tool calls over many turns, which is what a coding agent runs on.

Recursive’s guide uses OpenCode, the open-source coding agent: ollama launch opencode, then choose Flint. For a manual setup, it recommends Flint’s defaults (temperature 1.0, top-p 0.95, top-k 20), an output budget of 65,536 tokens, and leaving the reasoning-effort setting out, because Flint always reasons. Ollama’s integration recommends at least 64K of context.

Enjoy runs OpenCode beside Claude Code, Codex and the other labs’ agents, and its model picker lists every model OpenCode has set up, local ones included. That makes Flint useful in two ways:

• When a plan runs low. Switch the conversation to OpenCode with Flint, and the work carries on without a limit or a bill.
• On the night shift. Save routine upkeep as a workflow on a schedule, and Flint works through the night on your computer. Your code never leaves it. Enjoy runs scheduled work only while your computer is awake and Enjoy is open.

Keep a frontier model for the hard parts. Recursive’s own advice is to give Flint smaller, well-defined steps, and its card warns that it can still make “inappropriate tool calls,” so keep the agent’s permissions tight and review what it changes.