Unverified50% confidenceBenchmarkExact time
Via --quantization quark_mxfp4, SGLang dequantizes NVFP4 weights and requantizes them to MXFP4 at load time, achieving GSM8K accuracy recovery of 97.5%–100.2% on models such as MiniMax-M2.7 and GLM-5.1.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/5/2026
First Seen
Valid until: 12/4/2026
Sources
Related Entities
Related Claims
VerifiedGGUF量化格式（如Q4_K_M）通过将权重精度从32位浮点压缩到4位整数，通常只带来约1-5%的性能损失71% similarUnverifiedIn GGUF quantization naming, Q4_K_M denotes a 4-bit quantization, Medium-precision variant using the K-Quant grouped quantization method70% similarUnverifiedIn the Q4_K_XL scheme, MoE layers use MXFP4 and the remaining layers are compressed to Q8_0, achieving 96% same on top-1% prediction consistency69% similarUnverifiedQwen 2.5 7B使用4组GQA配置，推理时KV Cache体积相比MHA减少约75%67% similarUnverifiedQuantization methods such as GPTQ, AWQ, and SqueezeLLM can still maintain over 95% of the original model's performance at 4-bit precision.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/860553API
curl https://kongchang.com/api/v1/knowledge/claims/860553MCP
get_claim(id=860553)