Recursive Introduces Flint Model

Flint is a 40-billion-parameter mixture-of-experts coding model designed to run locally with high speed and strong agentic coding performance. It activates only 2.5 billion parameters per token and leads MoE models under 50B on benchmarks such as Terminal-Bench 2.1, SWE-Bench Pro v2, and LiveCodeBench, while matching dense 25-40B models at more than five times the throughput. Optimized inference achieves over 125 tokens/sec on an M4 Pro MacBook and over 150 tokens/sec on a 64GB NVIDIA RTX Spark at short contexts. Flint supports context windows up to 512k tokens (recommended within 256k), works offline to keep data local, and can require more explicit decomposition and instructions for complex tasks compared with larger closed-source coding models.

Flint was developed in under four months using an automated research system called Eureka Machine, which searched architectures, training recipes, quantization strategies, and hardware kernels. Key design choices include a 40B total / 2.5B active MoE architecture, aggressive 4-bit quantization (including the lm_head) recovered via an offline distillation stage, and specialized kernels for Apple Silicon and NVIDIA RTX Spark (including a Metal megakernel for M4/M5). Weights are available on Hugging Face under Apache 2.0; Flint is compatible with llama.cpp, Ollama, LM Studio, vLLM, and more, and Recursive also ships a lightweight rec inference engine optimized for M4/M5.