LFM3-1.7B quantization report: near-lossless at 3.6 GB VRAM, 2.2 GB with modest degradation
crusaderky · reddit · 2026-11-03
A detailed community quantization report on LiquidAI's LFM3-1.7B: per LiquidAI's own benchmarks it beats MiniCPM5-2B across the board while offering 2x the context and half the VRAM—about 3.6 GiB for a quasi-lossless quant vs 7.1 GiB for comparable models, and 2.2 GB with only modest degradation. The report maps the best quant levels on a KL-divergence-vs-VRAM curve, with interactive HTML plots and raw data published in the pixi-llm-recipes repo.
More from Models
- "Sonnet 5.5 used to be smarter": users sense capability regression — nlarusstone · 2026-11-03
- Apple Intelligence lands in China in December, powered by Alibaba's Qwen instead of ChatGPT — winer666 · 2026-11-03
- Only recent models seem to know "weight surgery" — are labs building Bayesian weight priors? — Tkaraletsos · 2026-11-03
- Victor Taelin floats labs replacing backprop with models editing their own weights at scale — arthurcolle · 2026-11-03
- Custom fine-tunes now beat frontier models at ~1/20th the cost, says investor — bindureddy · 2026-11-03
- OpenAI updates Model Spec: intent over literal instruction-following — sven_ai · 2026-11-03