Recursive ships Flint: a 40B MoE coding model (2.5B active) built in under 4 months by AI agents
AIFlow_ML · x · 2026-10-22
- Recursive (Richard Socher) released Flint, an Apache 2.0 MoE model for fast agentic coding on local hardware: 40B total params, 2.5B active per token, built from scratch in under four months by a small team plus the company's Eureka Machine automated research system.
- Leads sub-50B MoE models on SWE-Bench Pro v2, Terminal-Bench 2.1 and SWE-Bench Verified; rivals 25–40B dense models while running 5x+ faster.
- Over 125 tok/s on an M4 Pro MacBook Pro and 150+ tok/s on a 64GB RTX Spark at short contexts—up to 2x comparable models.
Related event: Recursive releases Flint, a fast local 40B MoE coding model(108 posts)→
More from coding & agent
- Hamel Husain and Shreya Shankar launch 'Evals for AI Engineers' book Nov 12 — HamelHusain · 2026-10-23
- Meta's MemoProBench: frontier models score ≤15% on memory-driven proactive assistance — RulinShao · 2026-10-23
- Agentic AI Foundation adds four projects from Anthropic, OpenAI, Hugging Face and Alibaba — rseroter · 2026-10-23
- Matt Pocock: Semantic diffusion has hit spec-driven development—how should we define SDD? — mattpocockuk · 2026-10-23
- Your agent can push anything you want to read straight to your Kobo ereader — lucasmeijer · 2026-10-23
- Thorsten Ball demos spinning up a new orb provider on an exe.dev VM with Docker — davidcrawshaw · 2026-10-23