Be The Investor ResearchRecursive's Flint Brings a 40-Billion-Parameter Coding Model to Laptops
Recursive released Flint, an open-weight coding model with 39.8 billion parameters but about 2.5 billion active per token. It runs locally and was built with automated research, though company benchmarks need independent validation.
Recursive has released Flint, a coding model designed to run locally on a well-equipped laptop rather than through a paid cloud service. The model has 39.8 billion total parameters, but its mixture-of-experts design activates about 2.5 billion for each token, reducing the computing needed for a response.
The open weights are available on Hugging Face under the Apache 2.0 license. Running the model locally can keep source code on the user's machine, support offline work and avoid usage charges tied to cloud-based tokens. It also shifts hardware, setup and maintenance requirements to the developer.
Recursive says a small team built Flint in less than four months with help from its automated research system, the Eureka Machine. The system proposed ideas, ran experiments and influenced architecture, training and inference. Company claims include low-bit training that reduced the model's memory footprint by 71% with little quality loss and custom kernels that improved speed.
On a MacBook Pro with an M4 Pro chip, Recursive reported roughly 131 generated tokens per second, versus 58 to 69 for selected comparable models. It also said Flint exceeded other sub-50-billion-parameter models on SWE-Bench Verified, which measures the ability to fix real software bugs. These are useful claims, but benchmark setup, quantization and hardware configuration can change results substantially.
The company itself says Flint may require more guidance than the largest paid coding assistants. Breaking large assignments into smaller steps can improve performance, which is a practical limitation for developers expecting autonomous execution.
Flint is also an experiment in AI-assisted AI development. That is not the same as uncontrolled recursive self-improvement. Humans supervised the research system, selected goals and retained the ability to stop the process. The launch demonstrates automation across parts of model development, not a self-directing system that independently improves without oversight.
For Apple, Nvidia and AMD, local models can support demand for capable consumer and workstation hardware. For cloud providers, the effect is less direct: some coding workloads may move to devices, while training and larger tasks still require substantial data-center capacity.
Independent reproduction will matter more than launch-day comparisons. Developers should evaluate Flint on their own repositories, measure error rates and confirm that local privacy requirements are met. If the model's speed and quality hold up, it could make private, offline coding assistance practical for a wider group of users.