Summary
Sam Witteveen looks at Flint, a new open-weights coding model from Recursive that aims to match mid-sized dense models at far higher speed. It is a mixture-of-experts model with 40 billion total parameters and about 2.5 billion active per token, using 256 experts with six selected per token plus one shared expert. It supports 262k context, always-on reasoning, and an Apache 2 license.
On a MacBook Pro with an M4, Recursive’s own engine reportedly reaches around 130 tokens per second, roughly ten times faster than the dense Qwen 3.8 27B at about 13. Benchmarks are mixed: Flint leads similar MoE models and ties Qwen 3.6 on LiveCodeBench, but trails the dense 27B on Terminal Bench (55 versus 73) and SWE-bench. Weights are available on Hugging Face in BF16, INT4, GGUF, and NVFP4 formats, with support in Ollama, LM Studio, llama.cpp, vLLM, and SGLang.
The video also explores Recursive itself, which emerged from stealth in May with $650 million at a $4.65 billion valuation, and its automated research system, the Eureka machine. Witteveen discusses the open questions about which models power it, and the lesson that evaluators become attack surfaces as optimization improves.
📺 Source: Sam Witteveen · Published October 23, 2026
🏷️ Format: News Analysis
