Hands-on: binary quantization is lossy for most models, ternary works fine
antoine_chaffin · x · 2026-10-01
- Antoine Chaffin (lightOn) shares hands-on quantization experience: binary quantization is noticeably lossy for models not trained for it (or without special tricks).
- He notes ternary quantization delivers pretty good results even with arbitrary models, making it a practical middle ground.
More from Models
- ChatGPT and Claude refuse scraping — unless you give them a nice reason — jdluk87 · 2026-10-01
- DeepSeek V4 groupchat memory revealed: it remembers memes and poignant moments — liminal_bardo · 2026-10-01
- Gemini 4 Argon launches with 1M-token output, tops 13 of 19 benchmarks at $2/$10 intro pricing — Latent Space · 2026-10-01
- Developers claim unreleased Gemini 4 Argon scores 91.9 atop VibeCodeBench — sunjiao123sun_ · 2026-10-01
- Claude Opus 5.5 paints Qi Baishi-style ink shrimp in live WebGL shaders — anselm · 2026-10-01
- Did Opus 5.5 enter a "nerfed" phase? LiveNerf baseline update fuels debate — Kazaan · 2026-10-01