Statistically-lossless Qwen3.8-27B released: up to 57% smaller weights with 98-99% EAR fidelity
udmrzn · x · 2026-10-23
Built on the SLQ method from ISTA's DASLab (the group behind GPTQ, GSQ and RCO), three statistically-lossless quantized Qwen3.8-27B checkpoints are out at 23.7–30.5 GB—up to 57% smaller than BF16.
- EAR (measured overlap with BF16's next-token probabilities, top-10 tokens renormalized) hits 98–99%, vs 97.71% for native FP8
- The search optimizes EAR directly, assigning more precision to sensitive projections while keeping activations in BF16
- Apache 2.0, works with vLLM + Humming; GGUF versions coming next week
More from Models
- OpenAI blocks manual reset with 13% usage left: 'no reset needed right now' — AccomplishedSugar490 · 2026-10-23
- Meta's 816B Muse Spark Open scores 39, losing to 180B Qwen3.8-Flash-Next at 40 — BTA_Labs · 2026-10-23
- Gemini refuses to generate image in Docs, blames document size limit — viksit · 2026-10-23
- Dev jokes his true calling is telling everyone to try the latest DeepSeek — intellectronica · 2026-10-23
- SubQ Snap 1.1 ties for first on Artificial Analysis Decision Index with 384K context — ArtificialAnlys · 2026-10-23
- PolyAI unveils Dialog-RSN-1: audio-native voice model answering in under 300ms — matthen2 · 2026-10-23