New Qwen3.8 27B quants beat ISTA and Unsloth byte-for-byte on three corpora
Dutchnamn · reddit · 2026-09-23
A Redditor released AP-GGUF quants of Qwen3.8 27B that beat ISTA and Unsloth quants by a wide margin byte-for-byte on three corpora (KLD and top-1% tested 3x). Generated over a week of continuous GPU/CPU time on a single Strix Halo; available on Hugging Face.
More from Infra
- Rivet launches SQLite Forking: 20ms agent-thread forks with zero data copying — samgoodwin89 · 2026-11-11
- French netizen proposes datacenters pay local residents compensation — clemnt · 2026-11-11
- GPU Rental Prices Hit Record Highs: B200 at $6.97/hr, H100 at $3.06 — inductionheads · 2026-11-11
- A Datacenter in Every Office: Building Work-Safe Local AI Racks — dee_hw · 2026-11-11
- "Compute is about power": the AI race's real bottleneck is energy, not chips — _sholtodouglas · 2026-11-11
- CoreWeave CEO says he canceled his 10b5-1 plan, halting insider selling — firstadopter · 2026-11-11