Recursive Flint: open 40B coding model runs on a laptop

AI startup Recursive released Flint, its first model, on October 21, 2026. Flint is an open-weight coding model built to run on a developer's own machine, and its weights are free to use under the Apache 2.0 license. Recursive says it runs at more than 125 tokens per second on a MacBook Pro, fast enough to power a coding agent without sending code to a cloud service.

Open-weight means anyone can download the model's trained numbers and run it. A token is a small chunk of text, often a short word or part of one. Richard Socher introduced the model on X, RuntimeWire reports.

How Flint is built

Flint is a mixture-of-experts (MoE) model. Instead of one large network, it holds many smaller "expert" networks and runs only a few of them for each token. That makes it much faster than its total size suggests.

According to its Hugging Face model card, Flint has 39.8 billion parameters in total, which Recursive rounds to 40 billion. About 2.5 billion are active for each token. Each of its 47 MoE layers holds 256 experts, picks 6 per token, and also runs one shared expert every time.

Spec	Flint
Total parameters	39.8 billion
Active per token	About 2.5 billion
Layers	49 (2 dense, 47 MoE)
Context window	262,144 tokens per the model card
Full download	About 80 GB (16-bit weights)
Quantized download	23GB on Ollama
License	Apache 2.0

The sources differ on context length, the amount of text the model can consider at once. The model card, LM Studio and Ollama list about 262,000 tokens. Recursive's announcement says up to 512,000, with 256,000 recommended.

Recursive says a small team built Flint from scratch in under four months. It credits its own automated research system, the Eureka Machine, with work across "nearly every part" of the build. That includes the architecture search, the quantization and the inference code. Quantization shrinks a model by storing its numbers with fewer bits. Recursive says its 4-bit NVFP4 version scores within 0.7 points of the full 16-bit model.

Benchmarks and speed

Recursive reports these scores on its model card, as the average first-try success over at least five attempts:

• SWE-bench Pro v2: 89.7%
• SWE-bench Verified: 74.2%, a test built from real GitHub bug reports
• Terminal-Bench 2.1: 54.9%, tasks done in a command-line terminal

These are the company's own numbers, and the card notes that "evaluation setups and runtimes can affect scores." Recursive claims Flint leads mixture-of-experts models under 50 billion parameters on agent coding tests. It also says Flint rivals dense 25-40 billion parameter models while running more than five times faster.

Speed depends on hardware and context length. Recursive cites more than 125 tokens per second on an M4 Pro MacBook Pro. It cites more than 150 on a 64GB Nvidia RTX Spark at short contexts. OpenHands, which builds an open-source coding agent, measured 100 or more tokens per second on the RTX Spark at a 120,000-token context.

Where to run it

Flint ships through the usual local tools: Ollama, LM Studio, llama.cpp, Unsloth, vLLM and SGLang. Recursive also released its own inference engine, called rec, tuned for M4 Pro and M5 Pro and Max chips. Ollama's listing includes launch commands for Claude Code, OpenCode, Hermes Agent and OpenClaw. Ollama already lets Claude Desktop use local open models, so Flint slots into that setup too.

OpenHands supported Flint from launch day. "Recursive's new model shows real promise for coding and agentic workloads," said Graham Neubig, its co-founder and chief scientist. The model is also on OpenRouter, a service that serves many models through one API, though its page says Flint is not yet available through EU in-region routing.

Crypto Briefing reports Recursive has raised $650 million at a $4.65 billion valuation. No other source we read confirms those figures.

What this means for developers

Flint is worth a test if you want a coding agent that keeps your code on your own machine. The Apache 2.0 license allows commercial use and fine-tuning, with no usage fees for the weights.

Check your memory first. LM Studio lists 24GB of system memory as the minimum, and the quantized Ollama build is a 23GB download. Crypto Briefing reports agent use is workable on machines with 48GB. Serving the full 16-bit model with SGLang needs one B300 or H200 data-center GPU per copy, per the model card.

Run it on your own repository before trusting the benchmarks. Use Recursive's suggested settings, temperature 1.0, top-p 0.95 and top-k 20. Give it room to think, too. The card warns that long reasoning "can exhaust the output budget before an answer is complete," and suggests at least 32,768 output tokens for hard problems.

Test long contexts yourself, given the disagreement over the limit. EU teams with data-residency rules should run it locally rather than through OpenRouter for now.