RTX 3060 Test: ByteShape Ultra-Low-Bit Quant Far Behind GSQ Despite Paper Similarity
zyxciss · reddit · 2026-09-23
On an RTX 3060 12GB + 16GB single-channel DDR4 (CachyOS, llama.cpp), the author compared two ultra-low-bit quants of Qwen 3.8 27B:
- ISTA-DASLab GSQ-RCO-IQ3-XXS (10.4GB, 2.5 BPW, MTP enabled): 29 tok/s at full context, 34-40 tok/s lower;
- ByteShape IQ3-XXS: 500MB smaller, advertised as extremely close to BF16 by KL-divergence.
In practice the gap was large: GSQ generated a 3D voxel diorama in one shot under 55k tokens, while ByteShape took 3 attempts and 98k+ tokens without finishing; web dev results also fell short of the advertised benchmarks. The takeaway: paper similarity metrics for low-bit quants don't guarantee real task performance. The author also asks for model suggestions for RTX 3060.
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