Recursive released code and demos enabling an eight agent research swarm to run locally on a single 48GB M4 Pro Mac using one shared instance of its 40B parameter Flint model. Given a problem and a target score, the agents write code, test ideas and build on successful experiments directly on the device without external compute.

The setup builds on Recursive’s Eureka Machine, an automated AI research system that helped build Flint from scratch in under four months across pre-training, post-training, reinforcement learning and inference. Co-founder Tim Shi said the startup “actually didn’t have any pre-training, post-training effort until like 1 month ago,” relying on the automated system to pick the mixture of experts architecture with 2.5B active parameters and to quantize the language model head to 4 bits, a move usually avoided by human researchers that stayed within 0.7 points of BF16 precision.

Released under an Apache 2.0 license with a 262K token context window, Flint posted 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified, placing it just behind the more than 10x larger Inkling model. The model generates more than 125 tokens per second on an M4 Pro MacBook Pro and over 150 tokens per second on a 64GB Nvidia RTX Spark, while independent tests clocked it 4.7x faster than Nemotron 3.5 in local agent orchestration and nearly 4x faster than Qwen3.8 27B on a Mac Studio at 110 tokens per second versus 28.