Show HN: 8-bit Qwen3.8-27B decodes 1.7x faster than BF16, slower at 16K context

(github.com) by gltanaka | Aug 19, 2026 | 0 comments on HN