Recursive open-sources Flint, a fast MoE coding model auto-built by its Eureka Machine in 4 months
Sam Witteveen · youtube · 2026-10-23
Sam Witteveen reviews Flint, the first public model from Recursive AI — a fast mixture-of-experts coding model built to take on Qwen3.8-class models, billed as the first open RSI (recursive self-improvement) coding model.
Key points
- Flint was auto-built by Recursive's "Eureka Machine" in under 4 months, including architecture search
- Open weights in four official versions, plus NVFP4/GGUF quantizations on Hugging Face
- Benchmarked against Qwen 3.6, Laguna XS, and Qwen 3.8 27B, with a speed-vs-score comparison highlighting its low-latency focus
- Runs locally via Ollama, LM Studio, llama.cpp, and vLLM; demo includes NVFP4 on an RTX PRO 6000, agent tasks and tool use
The video also explains how the Eureka Machine works and which GPUs can run Flint.
More from coding & agent
- FilmFlow: Multi-agent framework turns natural language into 3D camera trajectories — Yiming Zhang · 2026-10-23
- Raw-RAG beats complex agentic memory systems on 6 of 7 long-term memory benchmarks — omarsar0 · 2026-10-23
- Discussion: Serving Skills Over MCP to Close the Agent Learning Loop — curious_vii · 2026-10-23
- Proposal: serve skills over MCP, compress agent traces into merged skill updates — curious_vii · 2026-10-23
- Open-source MCP app renders 3D anatomy inline in Claude, plus an Obsidian plugin — pitfa19 · 2026-10-23
- Hooking an AI agent to a local Obsidian vault on a Mac mini: what worked and what didn't — techlatest_net · 2026-10-23