Unverified50% confidenceBenchmarkExact time
In Red Hat's research, GPTQ and AWQ 4-bit methods with the same group size scored 67 and 63 respectively on HumanEval.
1
Sources
50%
Confidence
Long-term
Relevance
9/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAWQ (Activation-aware Weight Quantization) typically outperforms GPTQ at 4-bit quantization.69% similarUnverifiedGPTQ（2022年）和AWQ（2023年）能将70B参数的LLM量化到INT4精度，困惑度损失不到1%63% similarUnverifiedQuantization methods such as GPTQ, AWQ, and SqueezeLLM can still maintain over 95% of the original model's performance at 4-bit precision.61% similarUnverifiedReddit社区的大规模基准测试基于BeeLlama.cpp v0.4.0，在Qwen 3.6 27B和Gemma 4 31B两款模型上累计跑了413组配置对比，其中Qwen模型238组、Gemma模型175组60% similarUnverified在GPT-4发布初期，主流模型在GPQA上的准确率约为35-40%，人类领域专家基准线约为65%，随机猜测为25%58% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/925956API
curl https://kongchang.com/api/v1/knowledge/claims/925956MCP
get_claim(id=925956)