LFM3 1.7B quantized with RCO/GSQ beats stock GGUFs; non-English tasks degrade far more
maximelabonne · x · 2026-11-04
bnjmnmarie used RCO and GSQ quantization to produce improved 1GB GGUF checkpoints of LiquidAI's LFM3 1.7B, with RCO directly optimizing EAR as its objective (scores computed on a held-out dataset). A benchmark chart shows the new GGUFs consistently outperform similarly sized donor GGUFs from LiquidAI and AtomicChat on EAR@10/EAR@100.
The author also measured EAR in French, Japanese, and Hindi, finding quantization causes noticeably larger degradation on non-English tasks—a commonly overlooked pitfall. Checkpoints and results are open on Hugging Face.
More from Models
- Leaker hints: no "Fable 5.5" or "GPT-6.2 Astra" today — but Thursday is the date — daniel_mac8 · 2026-11-04
- Trajectory launches post-training platform: supply tasks, data and eval criteria, deploy trillion-parameter models — udmrzn · 2026-11-04
- GPT-6.2 spotted: New OpenAI model rumors point to "astra" incoming — udmrzn · 2026-11-04
- Larger tokenizers suffer more multi-epoch regret, small vocabs survive repeats — SonglinYang4 · 2026-11-04
- Anime image forensics tool foidfinder detects character mashups and booru tags — AlpinDale · 2026-11-04
- Sonnet reportedly behaves with a different personality on Medium vs Max tiers — AaronBergman18 · 2026-11-04