Mach 2 Medium sets new Pareto frontier for Qwen 3.8 compression: 1.5% loss at 26.7GB, beating Unsloth
pbaylies · x · 2026-11-05
@syzygyeng's Mach 2 Medium compression of Qwen 3.8 Flash-Next sets a new Pareto frontier across GPQA Diamond, Terminal-Bench 2.1, DeepSWE v1.1, and AutomationBench: 26.7GB of weights with only 1.5% average loss vs BF16 — 10.9GB smaller than Unsloth's comparable UD-IQ1S and well ahead of GSQ (ISTA) tiers. Higher-fidelity Unsloth quants (UD-IQ2XS/XXS) remain lossless but need 50GB+ VRAM.
More from Infra
- Nationwide voters rejected data centers in 40+ provisions as AI's scientific wins go unnoticed, argues investor — beffjezos · 2026-11-05
- CMU paper unifies pretraining, SFT and RL into one likelihood-based algorithm — burny_tech · 2026-11-05
- Thousands of GPUs sit idle while frontier labs scramble for compute — _arohan_ · 2026-11-05
- Ontario's Ford Government to Approve AI Data Centres 'Sooner Rather Than Later' — shadowt1tan · 2026-11-05
- PyTorch Day Japan agenda: backpropagation for real-time robot inference with ExecuTorch — PyTorch · 2026-11-05
- VibeSys uses agents to auto-generate Qwen3.8 serving engines for AWS Trainium2 — bariskasikci · 2026-11-05