ExLlamaV3 update: MoE expert CPU offload, GLM-5.3-Flash, self-calibrated quants
Unstable_Llama · reddit · 2026-09-01
turboderp's local inference engine ExLlamaV3 shipped a batch of major updates:
- CPU offload of MoE experts for tight VRAM budgets
- ngram disk offload support for Qwen3.8-Flash-Next, plus GLM-5.3-Flash in exl3 format
- A new self-calibrated optimization quantization technique that auto-tunes per model
- Countless other optimizations
The author notes NVIDIA users who haven't tried it lately are missing out; the attached cat SVG was generated with Qwen3.8-Flash-Next.
More from Infra
- SGLang Summit sponsors grow from 11 to 36 in one week; Lip-Bu Tan, Lilian Weng to speak — BanghuaZ · 2026-10-18
- Yacine jokes his datacenter KPI is gigawatt-hours: get paged if not burning enough power — yacineMTB · 2026-10-18
- 1.68x inference speedup drops, directly accelerating GRPO post-training — yacineMTB · 2026-10-18
- OpenRouter is only a fraction of the total addressable market, poster argues — sloppenheimer · 2026-10-18
- Developer builds change alerts for LLM prices and retirements across Bedrock, Azure and Vertex — jsandhu0780 · 2026-10-18
- 12.3M tokens in 27 minutes for $0.14: agent session with 99.6% cache hit — teortaxesTex · 2026-10-18