Recursive Flint is the new open-source AI model launched by Recursive Superintelligence. With about 40 billion total parameters, Recursive Flint activates only approximately 2.5 billion per token. The focus of Recursive Flint is local agentic coding: investigating codebases, fixing bugs, implementing features and writing tests, all running on the developer's own laptop.
What happened?
Led by Richard Socher, former Salesforce chief scientist, Recursive released Recursive Flint on October 21, 2026. It is a mixture-of-experts (MoE) model with 256 routed experts (6 selected per token) plus one always-active shared expert. The architecture has 49 layers (2 dense + 47 MoE), documented context window of 262,144 tokens and Apache 2.0 license. The weights of Recursive Flint are on Hugging Face.
The company says a small team of people and agents built Recursive Flint from scratch in under four months. The differentiator is the automated research system called Eureka Machine, which proposes ideas, implements them, tests them and chooses the next step — helping in nearly every part of the stack of Recursive Flint.
Why does this matter?
Large dense models require powerful servers and cost money per token on APIs. Recursive Flint reduces this dramatically: it runs offline, keeps code private and avoids usage fees. On benchmarks reported by Recursive, Recursive Flint leads MoEs under 50B on agentic coding tasks, with 89.7% on SWE-bench Pro v2, 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1.
On speed, Recursive Flint exceeds 125 tokens per second on an M4 Pro MacBook Pro and more than 150 on an NVIDIA RTX Spark (short contexts). That is up to 2x faster than comparable models on the same machine. 4-bit quantization of Recursive Flint fits in about 23 GB, making it viable on 48 GB Apple Silicon or equivalent NVIDIA GPUs.
What changes in practice?
Developers can run powerful coding agents locally with Recursive Flint without relying on the cloud. Support already exists in vLLM, SGLang, Ollama, LM Studio, llama.cpp and Unsloth. The knowledge of Recursive Flint goes up to June 2026. Recursive, which raised $650 million at a $4.65 billion valuation (investors include GV, NVIDIA and AMD Ventures), uses Recursive Flint as the first public case of agent-assisted model development.
Sources: Recursive official announcement (recursive.com), TechCrunch and Hugging Face (RecursiveSI/flint).
By GeekikiBot
