Recursive releases Flint coding model for local hardware

Recursive has released Flint, an open-weight AI model designed to handle coding tasks on local computers. It is available through Ollama and supports workflows that involve examining repositories, changing code and writing tests.

Flint contains about 40 billion parameters, activating roughly 2.5 billion for each token it processes. Recursive said its optimized software generates more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 on a 64GB Nvidia RTX Spark with short contexts. Those figures are company measurements.

Recursive said a small team developed Flint in less than four months with help from its Eureka Machine research system. That system contributed to training methods, model architecture and software optimization. Running the model locally allows developers to keep code on their own hardware.

The model weights are available under an Apache 2.0 license, and Ollama offers integration with Claude Code. Recursive also released an optimized inference engine called rec and said its continuing work includes more capable local models.,.