Recursive.SI's Eureka Machine builds Flint-40B-A2B from scratch in 4 months, fast local coding
ChengleiSi · x · 2026-10-22
Recursive.SI (founded by Richard Socher) released Flint-40B-A2B, a fast local coding model built by its automated research system, the Eureka Machine.
- Strong scores: 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2, 74.2% on SWE-bench Verified
- Trained fully from scratch (tokenizer → pretrain → midtrain → SFT → RL), with the complete training pipeline built in just 4 months
- Up to 2x faster than comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark
- Open weights on Hugging Face; runs locally or free on opencode and pidot.dev
More from Models
- You Can Judge a New LLM by Which Benchmarks Its Team Picks for the Announcement — mariofilhoml · 2026-10-22
- Jensen Huang Keynote Demo Shows 'GPT-6 Astra High' in the Model Picker — LauraModiano · 2026-10-22
- BaseLabs open-sources mammoth-1 and elephant-1, top-ranked decision models — saranormous · 2026-10-22
- Richard Socher teases new release: "intelligence per dollar just went up a lot" — altryne · 2026-10-22
- Four models quietly reproduced Sagrada Família's magic square in offline 3D scenes — ArtificialAnlys · 2026-10-22
- French users urged to report Google as promo AI Pro subs lose Gemini 4 Argon access — Dazzling_Drummer_531 · 2026-10-22