Introducing Flint

gist

Recursive releases Flint, an Apache 2.0 open-weight coding MoE with 40B total and 2.5B active parameters, built for fast agentic coding on local hardware. Recursive reports 54.9 on Terminal-Bench 2.1, 89.7 on SWE-bench Pro v2, and 74.2 on SWE-bench Verified, and 131.22 tokens a second at 1K context on a 48 GB M4 Pro through MLX, about twice Qwen3.6 35B A3B's 65.55.

ideas

• Fewer active parameters buy local speed. Flint activates about 2.5B of 40B parameters per token, so it decodes more than 125 tokens a second on an M4 Pro MacBook Pro and more than 150 on an RTX Spark at short contexts, over five times faster than 25–40B dense models with similar coding scores.
• It leads its size class but not every dense model. Flint tops sub-50B MoE peers (Qwen3.5, Qwen3.6, Laguna XS 2.1) on the agentic benchmarks, while Qwen3.8 27B dense still scores higher, for example 73.0 against 54.9 on Terminal-Bench 2.1, at a fraction of the speed (13.01 tokens a second).
• An automated research system did much of the building. Recursive says its Eureka Machine contributed to data, optimizers, training recipes, architecture search, quantization, and kernels, letting a small team build Flint from scratch in under four months.
• Quantization broke a common rule. The Eureka Machine quantized the output layer to 4 bits and recovered accuracy with self-distillation on a coreset, leaving the NVFP4 model within 0.7 points of 16-bit; on M4 and M5 chips it produced a Metal megakernel fusing the whole decode pass.
• Weights and runtimes ship together. Weights are on Hugging Face under Apache 2.0, with support in llama.cpp, Ollama, Unsloth, LM Studio, vLLM, and SGLang, plus Recursive's own rec engine for M4 Pro and M5 Pro/Max; context reaches 512K tokens, with 256K recommended.

quotes