Recursive's Flint: a 40B coding MoE model that runs on a laptop

Recursive released Flint on October 21 under Apache 2.0: a 39.8B-parameter MoE model activating about 2.5B parameters per token. It claims over 125 tokens/s on an M4 Pro MacBook Pro and 150 tokens/s on a 64GB Nvidia RTX Spark, with only self-reported benchmarks so far.

• 39.8B MoE parameters, ~2.5B activated per token
• 49 layers: 2 dense and 47 MoE with 256 experts, 6 active
• Speed: 125+ tokens/s on M4 Pro, 150+ on RTX Spark 64GB
• Quantized files: NVFP4 ~22.9GB and int4 GGUF ~26GB