Recursive's 40B open model Flint ran 732 auto-experiments to improve NanoChat, nearing Claude Haiku 4.5
cong_ml · x · 2026-10-22
Recursive released early results from its automated AI research system, powered by Flint, a new 40B open model (2.5B active params/token) that runs on a laptop. Across 732 NanoChat experiments—each needing only 5 minutes on one H100—the system closed the loop of proposing ideas, implementing code, training, validating, and picking the next experiment, with reward-hack and variance checks before accepting progress.
Key points:
- The system claims SOTA on three benchmarks: fixed-budget LM training, small-model training speed, and GPU kernel optimization.
- Flint mostly tuned hyperparameters to a tight compute budget: smaller batches (524k→262k tokens), head LR 0.004→0.02, a deeper 9-layer lineage, nonzero final LR, and head refinement cut validation BPB from 0.9964 to 0.9797.
- That lands close to historical Claude Haiku 4.5 (0.9770) at the same budget, though Sonnet 4.6 (0.9744) is still ahead.
- All run artifacts are open-sourced for inspection and reuse.
More from Models
- Flint int4 hits 122.6 tok/s on Mac mini matching Qwen3.6-35B quality at 3.5x speed — _rockt · 2026-10-22
- Bayesian take: LLM benchmark noise means all models are probably tied — remilouf · 2026-10-22
- OpenAI's New Pricing Slammed for Zero Transparency as $500 Tier Reveals No Usage Numbers — Effective_Tap_9786 · 2026-10-22
- OpenAI's pricing pages never state actual usage limits, power users cry foul — Monstermage · 2026-10-22
- Nemotron 3.5 Super VL: NVFP4 on one Blackwell GPU beats BF16 on two, ~2.2x per-GPU throughput — TheZachMueller · 2026-10-22
- LFM2.5-24B-A2B Tested and Found Lacking: Looping Outputs, Broken Code, Full Configs Posted — tracagnotto · 2026-10-22