Qwen Image 2.1 quantized to 2x speed — but INT4 quantization loses to FP4
mesmerlord · reddit · 2026-09-22
A developer quantized the new Qwen Image 2.1 to replicate their paid tool's serverless setup (previously Flux klein + nunchaku), publishing weights as Mesmer-Image-21-Nunchaku on Hugging Face.
Key findings:
- INT8 conversion ballooned to 9GB; INT4/FP4 come in around 4.8GB
- Roughly 2x faster, but with quality hits: text survives well, background faces degrade (especially at 40 steps — 25 steps fares better), and diagrams suffer noticeably
- Full comparison benchmarks (BF16, FP4 for RTX 50xx, INT4 for 40xx and below) at mesmer.tools/benchmarks/qwen-image-2-1
- Surprising result: INT4 is worse than FP4, so 50xx-card users should take the FP4 route
- Author notes anyone with a coding agent could reproduce the same conversion pipeline themselves
More from Infra
- Gas-for-AI-datacenters forum co-located with CoRL shows energy demand spillover — carlosdponx · 2026-11-12
- a16z podcast: Mike Volpi on Cisco's 70+ acquisition playbook and the AI buildout — a16z · 2026-11-12
- AMD to pass Intel in server revenue by mid-2027 as share gains 6 points a year — BenBajarin · 2026-11-12
- Study: Reflective Data Center Roofs Actually Cool Daytime Ground by 1.8-2.2°C — SumitGup · 2026-11-12
- Tsinghua's BaGuaLu-2 tames MoE expert-load spikes at runtime, accepted to SC26 — jiqizhixin · 2026-11-12
- antirez backs low-bit DeepSeek 4.1 Flash over Qwen 27B dense on Apple M5 Macs — antirez · 2026-11-12