Flint
Flint is an open-weight coding model released in October 2026 by Recursive, the startup of Richard Socher, its cofounder and CEO. It has 39.8 billion parameters in total but activates only about 2.5 billion per token (a mixture-of-experts design), which makes it light enough to run on a laptop. Weights are under the Apache 2.0 license on Hugging Face, with a documented context of 262,144 tokens and lighter builds planned for llama.cpp, Ollama and LM Studio. The publisher says it built the model in under four months with its automated research system, which handled almost the whole chain, from data to the compute code for Apple and Nvidia chips.
Strengths
- Mixture of experts: 39.8 billion parameters in total, about 2.5 billion active per token, so far less compute per word than a dense model of the same size
- Apache 2.0 weights on Hugging Face, lighter builds planned for llama.cpp, Ollama and LM Studio: you can test it without a publisher account
- The publisher says plainly what falls short: the model needs more guidance than closed coding assistants, hence tasks split into steps
Limitations
- Scores and speeds all come from the publisher (tested on its own engine, short context), no independent measurement published so far
- The published checkpoint weighs about 80 GB, and the card gives no memory figure for the lighter builds: the video's "25 GB" is not confirmed
- The documented context is 262,144 tokens, the launch post says 512,000: stick to the model card, and know that accuracy is not guaranteed uniform across the whole window
Best for
- Developers on a recent Mac who want to experiment with a local coding model before deciding whether local is worth it
- Teams following the model race who want a concrete example of a model built with heavy automation