Recursive launches Flint: 40B MoE local coding model built in 4 months by its automated research system
RichardSocher · x · 2026-10-22
Recursive, Richard Socher's startup, released open-weight Flint, a mixture-of-experts model for fast agentic coding on local hardware.
- 40B total parameters with only 2.5B active per token; built from scratch by a small team in under 4 months
- Leads MoE models under 50B on agentic coding benchmarks: 89.7 on SWE-Bench Pro v2, 74.2 on SWE-Bench Verified, 54.9 on Terminal-Bench 2.1
- Speed: 125+ tokens/s on an M4 Pro MacBook Pro (131 via MLX) and 150+ tokens/s on a 64GB RTX Spark — roughly 2x faster than comparable sparse models and 5x+ faster than similarly scoring dense models; fully offline and private
Flint also validates Recursive's Eureka Machine, an automated AI research system running the full loop of ideation, implementation and validation. It contributed a sparser model shape, low-bit training that shrank the model 71% with minimal quality loss, custom kernels, and inference optimizations. Weights are on Hugging Face, runnable via ollama, llama.cpp, vLLM and more.
More from coding & agent
- Prime Intellect lays out continual learning roadmap: mining agent exhaust for self-improving models — willcb · 2026-10-23
- Claude spent 2 days reimplementing vLLM — readable code review reveals a maintainability nightmare — bingxu_ · 2026-10-23
- Percy Liang teaches new MIT course on self-evolving AI agents across four evolution levels — pliang279 · 2026-10-23
- Using GrokBot as an OSINT Tool to Trace the Origin of Forged AfD Emails — FlorianGallwitz · 2026-10-23
- intentoo: an MCP server where AI agents negotiate matches before humans ever meet — Appropriate-Ease2981 · 2026-10-23
- Two frontier labs quietly killed the temperature knob, breaking production determinism assumptions — SmartEnoughToFail · 2026-10-23