Qwen3.8-Flash-Next runs with full experts on just 37GB RAM
EyalToledano · x · 2026-08-29
The author shared a breakthrough in MoE inference optimization for Qwen3.8-Flash-Next. By storing 60% of the experts on disk and streaming them to memory on-demand (similar to n-gram streaming), the model achieves 40 tok/s decoding on a M4 Max with only 37GB RAM while using full experts (no pruning).
Related event: Optimization Lets Qwen3.8-Flash-Next Run on 37GB VRAM(4 posts)→
More from Infra
- Chip equipment beats bull case: 2026 WFE seen at $153B, $278B by 2030 — BenBajarin · 2026-10-16
- Steam now launches in 5.23s, beating Epic's 5.47s — and vLLM gets roasted for slow startup — AlpinDale · 2026-10-16
- Shadeform to Announce Six New Features on Oct 19: Lower GPU Costs, More Suppliers — _jaydeepkarale · 2026-10-16
- Fastino launches Training API for SFT/RL, claiming 25-85% cost cut vs Tinker — kalyan_kpl · 2026-10-16
- Dallas Fed: AI data center debt issuance equals ~1/8 of Treasury duration supply — _onionesque · 2026-10-16
- Running Wan2.1 14B FP16 on a 12GB laptop GPU: VRAM swap stress test finishes in 4m11s — RYuu-1120 · 2026-10-16