ExLlamaV3 3bpw Quantization Delivers Impressive Local Flash Performance

Users report that ExLlamaV3's 3bpw quantization runs Flash models at 1500 prefill and 29 tps on a single RTX 5090, with another achieving Q8-like quality on 3x RTX 3090s.

2026-09-20 ~ 2026-09-21 · 2 related posts

• Qwen 3.8 Next Flash at 3.05bpw EXL3 runs like Q8 on 3x RTX 3090s, dev reports — nicholas_the_furious · 2026-09-20
• ExLlamaV3 3bpw local Flash benchmarks: 1500 tps prefill on a single RTX 5090 — youcloudsofdoom · 2026-09-21