Unverified70% confidenceBenchmarkExact time
Re-quantizing INT4 weights to NF4 subjects each weight to two rounds of rounding, and official measurements show this increases the perplexity of Qwen3-8B by 1.6%
2
Sources
70%
Confidence
Long-term
Relevance
10/1/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedQ4 is 4-bit quantization and IQ3 is 3-bit importance-aware quantization; Qwen3 27B is approximately 11–12 GB in size under IQ3 and approximately 15–16 GB under Q4.70% similarUnverifiedAt a throughput of 259k, Qwen3.5 achieves approximately a 19× speed improvement, with hybrid attention contributing approximately 5×, MoE sparsification approximately 2×, and Multi-Token Prediction approximately 1.5×.67% similarUnverifiedINT4量化将原本由65536个不同值（FP16）表示的权重压缩到仅16个离散级别67% similarUnverifiedQ4_K_M采用分组量化策略，将权重张量划分为64或256维子块并独立计算缩放因子与零点，能将困惑度损失控制在1–3%以内67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/964985API
curl https://kongchang.com/api/v1/knowledge/claims/964985MCP
get_claim(id=964985)