Recursive releases its first model Flint: local coding MoE, 125+ tokens/s on a MacBook

Flint is the first model from @Recursive_SI: a 40-billion-parameter MoE that activates only 2.5 billion parameters per token, purpose-built for agentic coding on local computers, with weights open-sourced under Apache 2.0. Its real highlight is not just the model itself, but that almost every stage was completed with the involvement of the company's automated research system, Eureka Machine, serving as the first real-world test of "AI doing AI R&D itself."

Model information · Positioning: the strongest agentic coding capability among MoE models with fewer than 50 billion total parameters (SWE-Bench Pro v2, Terminal-Bench 2.1, SWE-Bench Verified); coding scores are comparable to dense models with 25 to 40 billion parameters, while being more than 5 times faster. · Speed: in short contexts, over 125 tokens per second on an M4 Pro MacBook Pro and over 150 per second on a 64GB RTX Spark, up to 2 times that of comparable models. · Context: up to 512k, with the official recommendation to keep it within 256k. Ecosystem: supports llama.cpp, Ollama, LM Studio, vLLM, SGLang, and Unsloth, and additionally released rec, a lightweight inference engine optimized for M4 Pro and M5. · Official usage advice: suitable as an everyday local coding model, with data staying on the machine and no API fees. Complex tasks should be broken down and instructions written clearly; it needs more guidance than large closed-source models.

What Eureka Machine specifically did: it automatically runs the entire research loop: proposing ideas, implementing, testing, then deciding the next step based on results, exploring multiple directions in parallel, and combining effective ideas. There are three examples on Flint: 1. Architecture search: starting from a baseline with 35 billion total parameters and 3 billion activated, it searched configurations of expert count and width, and finally settled on 40 billion total parameters with 2.5 billion activated, achieving the best balance between speed and coding capability. 2. Unconventional quantization: normally the output layer (lm_head) is not quantized, but it quantized the output layer to 4 bit as well, because when few parameters are activated, the output layer accounts for a large share of the per-token read volume. To recover accuracy, it added an offline distillation step: using the 16-bit model as teacher, the quantized model was trained on carefully selected data sampled by itself, with the number of training steps nearly halved. In the end, the gap between the NVFP4 version and the 16-bit version was within 0.7 points. 3. Kernel optimization: dedicated kernels for low-bit projection, MoE, attention, and prefill generation; on M4 and M5, the entire decoding forward pass is fused into one large Metal kernel.