Recursive announced on Wednesday (October 21, 2026) Flint, its first public model: a mixture-of-experts (MoE) model with about 40 billion total parameters that activates only 2.5 billion per token, designed for fast agentic coding on local hardware.
According to the company, a small team of people and agents built the model from scratch in under four months. The Eureka Machine, Recursive’s automated research system, contributed to nearly every step — from data and training recipes to infrastructure, quantization and inference kernels.
What Flint offers
- Leading performance among MoE models under 50B on agentic benchmarks such as SWE-Bench Verified (74.2%), SWE-bench Pro v2 (89.7%) and Terminal-Bench 2.1 (54.9%), according to Recursive.
- Speed: over 125 tokens per second on an M4 Pro MacBook Pro and over 150 on NVIDIA RTX Spark at short contexts — up to twice as fast as comparable models.
- Open weights under Apache 2.0 on Hugging Face, with support for tools like llama.cpp, Ollama, vLLM and SGLang.
- Privacy: runs offline, without sending code to the cloud and without per-token fees.
The company recommends breaking complex tasks into subtasks, as the model, while capable, may need more guidance than large closed coding models.
Why it matters
Flint is presented as a practical test of recursive self-improvement: the idea that AI systems can help build better AI systems. Recursive, which emerged from stealth in May with $650 million in funding at a $4.65 billion valuation (backed by GV, Nvidia and AMD Ventures), bets that more capable models make the Eureka Machine a better researcher, accelerating the cycle.
In the current context of local models and agents, this reduces dependence on expensive APIs and keeps data private — relevant for developers and companies concerned about code leakage.
Sources
Transparency: This content was created, edited or reviewed with the aid of artificial intelligence. Information was cross-checked with public posts on X and sources available on the internet. Consult original sources for full context.
By GeekikiBot
