Qwen3.8-27B Quantization Analysis: FP8 vs 4-bit Loss
pbaylies · x · 2026-08-20
Spent 67 hours of model time benchmarking Qwen3.8-27B quantization on 4x RTX 3090s. Compared FP8, NVFP4, AWQ INT4, GGUF Q4KM, and NInfer across 4,800 tasks, 10,120 requests, and 14.5M reasoning tokens. The results were surprising.
More from Models
- Opus 5.5 value may end the habit of using the top model for everything — adonis_singh · 2026-10-16
- Peter Yang: Opus 5.5 Still Struggles to Hold Its Ground When Pushed Back — petergyang · 2026-10-16
- Robot Startup Ditches VLAs for Frontier LLMs in a Customer Deployment — ihorbeaver · 2026-10-16
- Musk amplifies user claim that Grok's computer use is fast enough to skip a Codex subscription — elonmusk · 2026-10-16
- PhoneLLM launches open voice-AI models with sub-100ms latency, tops PhoneBench at 83% — charles_irl · 2026-10-16
- Transluce vets: a year of neuron autointerp work still can't explain complex behaviors — aryaman2020 · 2026-10-16