Flint: A 40B Open Coding Model That an AI System Helped Build in Four Months
Recursive's Flint is an Apache 2.0 open-weight coding model that runs on laptops and was developed largely by the company's own AI research system, though the claims are unverified.
A startup called Recursive has released Flint, its first language model, built for coding agents and designed to run on your own hardware. The company says it took about four months and that an in-house AI system, the “Eureka Machine”, handled nearly every step of development: proposing ideas, testing them and refining the results.
Flint is a “mixture-of-experts” model with 40 billion parameters in total, but only about 2.5 billion are active for each word it produces. Think of a large team of specialists where only a few are called in for each question, which keeps it fast. Recursive says the system helped with training data, optimisers and the architecture, and wrote low-level code for Apple Silicon. On speed, the company reports over 125 tokens per second (roughly words per second) on an M4 Pro MacBook and over 150 on an Nvidia RTX Spark with 64 GB. The weights are on Hugging Face under the Apache 2.0 licence, which allows commercial use, and it supports context windows up to 512K tokens (how much text it can hold at once), with 256K recommended. It works with llama.cpp, Ollama, LM Studio, vLLM and SGLang.
Now the caveats. The benchmark claims, that Flint leads other sparse models under 50B parameters on coding tests such as SWE-Bench Pro v2 and Terminal-Bench 2.1, come from Recursive and have not been independently checked. Larger dense models like Qwen 3.8 still score higher on some tests. Recursive itself says Flint needs more guidance on complex tasks than big proprietary models, so it recommends splitting big jobs into smaller steps. And the article notes the release does not yet prove “recursive self-improvement”, which would mean showing that a model built by the Eureka Machine makes the Eureka Machine a better researcher.
What’s behind it: the dream behind the name is AI that improves AI, in a loop. Flint is an early data point: a small team using automated research to ship a competitive model quickly. Even without the loop, that lowers the cost of building decent models, which is good news for open and local AI.
What this means for you: if you are new to local AI, this is not a must-try; wait for independent reviews. If you already run models on a Mac or a gaming PC, Flint is a cheap experiment as a private coding helper, as long as you keep tasks small.
Sources
Source: https://the-decoder.com/recursive-shows-first-result-of-its-ai-driven-research-process-with-flint/
A Scam Detector That Runs on Your Phone: DeepMind and Singapore Fine-Tune Gemma
Google DeepMind and Singapore's GovTech are building an on-device model that warns users about scam chats, without sending conversations to the cloud.
