Socher: Flint was built in <4 months by an automated AI research loop, shrunk 71% via low-bit training
RichardSocher · x · 2026-10-22
Richard Socher reveals Flint was a real-world test of the "Eureka Machine," his system for automated AI research. It runs the full loop—ideation, implementation, validation—and uses results to decide what to try next. It was used in nearly every part of Flint's development, letting a very small team build the model in under 4 months.
Its contributions include a sparser model shape, low-bit training that shrank Flint by 71% with almost no quality loss, and custom kernels. Flint runs at 125+ tokens/sec on an M4 Pro MacBook Pro and 150+ tokens/sec on a 64GB NVIDIA RTX Spark—up to 2x faster than comparable models on short contexts and 5x+ faster than dense models with similar coding scores—fully local and offline.
More from Models
- Six months later, the best AI costs 20× less — Tgoc · 2026-10-22
- Single RTX 5080 runs Qwen3.8-Flash-Next at 150 tok/s decode, 4.4K tok/s prompt — yehyakar · 2026-10-22
- Pichai on Gemini 4 Argon: even Google can't get enough AI compute — sundarpichai · 2026-10-22
- Flint founder recalls OpenAI Codex debate: local interactive coding vs cloud agents — josh_tobin_ · 2026-10-22
- Recursive launches Flint: an open 40B MoE model built for local agentic coding — josh_tobin_ · 2026-10-22
- Recursive launches Flint: 40B MoE local coding model built in 4 months by its automated research system — RichardSocher · 2026-10-22