Recursive Superintelligence has released Flint, an open-weight mixture-of-experts coding model featuring 40 billion total parameters and 2.5 billion active parameters per token. Built using the company's automated Eureka Machine research system, the text-only model is designed specifically for coding agents to investigate codebases, fix bugs, implement features, and write tests. The model weights are publicly available on Hugging Face under the Apache 2.0 license.
The architecture comprises 49 layers, including 2 dense layers and 47 mixture-of-experts layers with 256 routed experts and one always-active shared expert. Flint supports a large context window of up to 512,000 tokens and is optimized to run locally on hardware such as 48 GB Apple Silicon Macs or NVIDIA GPUs using 4-bit quantization. On performance benchmarks, the model achieves 95.1 percent on IFEval and 74.2 percent on SWE-bench Verified, though it trails certain dense models on other tasks like Terminal-Bench 2.1.
This release matters because Flint offers high-speed local agentic coding capabilities, delivering over 125 tokens per second on consumer-grade hardware while matching the performance of much larger dense models.