Flint

Recursive Flint is a decoder-only mixture-of-experts model with 39.8B total parameters (approximately 2.5B active per token) designed for local agentic coding tasks including codebase exploration, bug fixing, feature implementation, and testing. It supports up to 512k context (recommended 256k), tool calling, and runs on consumer hardware such as 48GB Apple Silicon Macs or NVIDIA RTX Spark GPUs. The model was developed from scratch in under four months by a small team assisted by Recursive’s Eureka Machine automated research system, which contributed to architecture, quantization (NVFP4 within 0.7 points of BF16), and optimized inference kernels. Weights are released under Apache 2.0 on Hugging Face and integrate with llama.cpp, Ollama, vLLM, SGLang, and a custom rec inference engine.