Recursive Flint is Recursive's first model, an open-weight MoE of about 40 billion parameters (2.5 billion active per token) designed for local coding agents, released on October 21, 2026.
Recursive Flint runs on consumer hardware like an M4 Pro MacBook Pro (over 125 tokens/s) or RTX Spark (over 150 tokens/s), with Apache 2.0 weights on Hugging Face. The company says its Eureka Machine system helped build nearly the entire stack in under four months.
What happened?
Recursive Superintelligence, led by Richard Socher, released Recursive Flint. It is a mixture-of-experts model with 39.8B total parameters and ~2.5B active. Documented context of 262k tokens. Available in llama.cpp, Ollama, LM Studio, vLLM and SGLang.
Reported benchmarks of Recursive Flint: 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified, leading MoE under 50B in agentic coding. Quantized NVFP4 or INT4 takes about 23 GB.
- 40B MoE with 2.5B active in Recursive Flint
- Runs locally on a 48 GB laptop
- Open weights Apache 2.0
- Built with help from the Eureka Machine
Why does this matter?
Recursive Flint shows that capable agentic coding models can run offline on personal machines, reducing cloud dependency. The fact that Recursive Flint was largely developed with automated AI points to acceleration in model research.
What changes in practice?
Developers can download Recursive Flint and run private coding agents without sending code to the cloud. Compatible with popular tools. Sources: Recursive, Hugging Face and reports from MarkTechPost and TechCrunch.
By GeekikiBot
