Same quant, same 4090: why is Qwen 3.8 slower than 3.6 locally?
Specialist-2193 · reddit · 2026-08-19
A Reddit user asks why Qwen 3.8 runs slower than 3.6 locally despite identical quantization (Q4KXL + MTP), identical hardware (RTX 4090), latest llama.cpp build, and 150k context. With the same architecture and similar MTP prediction quality, throughput should match — the user wonders what besides thinking output length could cause the slowdown. A concrete local-inference performance debugging discussion.
More from Infra
- Phison CEO calls CXMT memory scare a 'ghost story', warns of materials shortage — BenBajarin · 2026-10-16
- Ben Bajarin: The Future of Compute Is Fungible as Hyperscalers Design for Workload Flexibility — BenBajarin · 2026-10-16
- AirLLM runs frontier-size models layer-by-layer: 671B on 12GB VRAM — Aiden_Tech_Ai · 2026-10-16
- NVIDIA Launches cuPotential in CUDA-X to Bring AI Agents to Semiconductor Materials Simulation — tomaszbednarz · 2026-10-16
- Q4-quantized coder model on 8-core CPU builds interactive solar system demo in one shot — IntrepidMindExplorer · 2026-10-16
- TokenSpeed MLA 2 ships for Vera Rubin: up to 3.44× faster Kimi K3 decode — zhyncs42 · 2026-10-16