Recursive Unveils Flint: A Laptop-Ready Coding Model Built by AI

Recursive, the AI startup led by former Salesforce chief scientist Richard Socher, has released its first model: Flint, a free coding assistant designed to run entirely on a developer's laptop. The 40-billion-parameter mixture-of-experts model activates only 2.5 billion parameters per task, making it lightweight enough for consumer hardware while delivering competitive performance on coding benchmarks.

Efficiency Meets Performance

In internal testing, Flint generates roughly 131 tokens per second on a MacBook Pro with an M4 Pro chip — about twice the throughput of comparable open models such as Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning. On Nvidia RTX hardware it exceeds 150 tokens per second. On SWE-Bench Verified, a benchmark measuring real-world bug-fixing ability, Flint outperforms all other models under 50 billion parameters, trailing only Alibaba's Qwen3.8 27B in a six-test average.

Why it matters for GPU / AI infrastructure: Flint demonstrates how architectural innovations — sparse mixture-of-experts design and low-bit quantization that shrinks the model by 71% with minimal quality loss — can shift inference workloads from centralized clouds to edge devices. For enterprises, this means reduced API costs, stronger data privacy, and the ability to run capable coding assistants offline. GPU cloud providers like AiGpu can complement this trend by offering hybrid workflows: heavy training in the cloud, lightweight inference on local GPUs.

Perhaps most striking is how Flint was built. Recursive's automated research system, dubbed the Eureka Machine, contributed to nearly every stage — from pre-training data selection to post-training optimization and inference tuning. A small team delivered the model in under four months, showcasing a practical step toward recursive self-improvement where AI accelerates AI development under human oversight. Flint is available now on Hugging Face under the permissive Apache 2.0 license.