Qwen 3.8 Next Flash at 3.05bpw EXL3 runs like Q8 on 3x RTX 3090s, dev reports
nicholas_the_furious · reddit · 2026-09-20
A local LLM enthusiast shares detailed benchmarks running Qwen 3.8 models on 3x RTX 3090s (250-265W, one on a TB4 eGPU):
- Daily driver: Qwen 3.8 27B UD-Q8KL on 2x 3090s in TP, FP16 KV cache, 250k+ context with no perceived degradation (vs 150k for 3.6), great as a coding agent.
- Qwen 3.8 Flash Next UD-Q4KXL produced a stunning Flappy Bird variant with full UI, animations, and moving pipes — far more world knowledge than the 27B — but RAM offload killed speed (35-50 t/s out, low-hundreds PP).
- The fix: EXL3 3.05bpw quant fits fully in VRAM with full context and FP16 KV cache, feels indistinguishable from Q8, and delivers 110+ t/s output and 1500+ t/s prefill.
Verdict: if you have 3x 3090s, run Qwen 3.8 Next Flash at exl3 3.05bpw; EXL3 isn't worth it for the 27B, and Unsloth Q3 quants may be worth comparing.
Related event: ExLlamaV3 3bpw Quantization Delivers Impressive Local Flash Performance(2 posts)→
More from Infra
- Show Reddit: ge, a Rust GGUF engine, benchmarks against llama.cpp on a 4GB RTX 3050 — Raptor2496 · 2026-11-12
- Musk: 100GW of space solar equals ~800GW on Earth, and ground solar needs batteries — elonmusk · 2026-11-12
- Real-world test: Kimi K3.1 Pro costs 5.4x less than Opus 5.5 per task despite 2x list-price gap — Arindam_1729 · 2026-11-12
- Dino10B: first 10-billion-scale nearest-neighbor search benchmark, 10.3TB of DINOv3 embeddings — philipvollet · 2026-11-12
- Many announced data centers may never get built, TechCrunch argues — TechCrunch AI · 2026-11-12
- Meta loses longtime infrastructure head Santosh Janardhan after 16 years — TechCrunch AI · 2026-11-12