Benchmarking Qwen3.8 27B Quantizations: 4-bit Holds Up, 1-bit Collapses
pmigdal · reddit · 2026-08-27
The author benchmarked various Unsloth quantizations of Qwen3.8 27B on popular benchmarks including FPQA Diamond, IFBench, and Terminal-Bench-2.1. Results indicate that the Q4KM quantization performs robustly and is sufficient for most tasks, while the 1-bit quantization version collapses or shows severe performance degradation.
More from Models
- Nat Lambert: wave of AI releases rumored — thinky, muse spark, Fable 5.5, OpenAI — natolambert · 2026-10-15
- Musk amplifies dev's verdict: Grok Bot easier to use than Codex and Claude — elonmusk · 2026-10-15
- Rumor: New Anthropic model imminent, apparently behind a batch of high-severity Chrome CVEs — matthew_d_green · 2026-10-15
- StartLux-Decision debuts RL-trained calibrated decision models, 27B scores 72.8 — startlux-models · 2026-10-15
- Meta's Muse Code slammed for harsh quotas: 21% left after 5 hours — ThomasScialom · 2026-10-15
- Claude Code error mentions 'Fable 5.1' and 'Opus 5', hinting at unreleased models — adonis_singh · 2026-10-15