LLM edits model weights to quantize: Nace quants hit up to 2x better KL divergence
m4rkGS · reddit · 2026-11-14
nace.ai released GGUF quantized weights of Qwen 9B/27B/35B-A3B/Flash-Next using Weight Space Programming: instead of brute-force hillclimbing tensor precision mixes (which plateaued at SOTA), frontier-model agents natively read and edit the smaller model's weights, proposing precision changes and keeping or reverting each edit based on measured KL divergence. Results at 128K context: 9B Q3KXL hits 90.5% top-1 agreement at 4.7GB, 35B Q3KS 91.2% at 15.3GB, 27B 97.7% at 17.3GB, Flash-Next 93.0%—leading their size class and up to 2x better KL than Unsloth/ByteShape/EXL3. Scaling to larger QFN/DeepSeek-style models is weaker; paper coming soon.
More from Models
- Arena weekly: Claude Fable 5.2 tops Agent Arena, Gemini 4 Hydrogen fastest — arena · 2026-11-14
- Open models trail frontier by six months, but that matters less than you think — latkins · 2026-11-14
- Japan's ai& launches Kimi K3.1 Pro and Meta's Muse Spark Open with day-0 pricing — DavidBennett__ · 2026-11-14
- Alan Thompson memo: 18 models in 13 days, Gemini 4 elements, RLI heading to 98% — dralandthompson · 2026-11-14
- Rumor: Anthropic Fable 5.2 shows unusual robotics capability jump as scaling beats specialized training — zephyr_z9 · 2026-11-14
- Anthropic rumored to drop massive multi-field science breakthroughs around its IPO — Dr_Singularity · 2026-11-14