Unverified50% confidenceFactExact time
In the Q8_K_XL scheme, MoE layers retain the original MXFP4, while the remaining layers that were originally FP8 are upcast to BF16, and this conversion is 100% lossless
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
8/6/2026
First Seen
Valid until: 11/4/2026
Sources
Q8_K_XL与MXFP4命名之争：FP8≠Q8_0的量化真相
Related Claims
Unverified开发者采取混合精度策略，将对质量敏感的层保留在BF16，仅对能从FP8获益的部分做转换61% similarUnverifiedUP主实测使用的是Unsloth的Q4_K_XL量化版本59% similarUnverifiedQLoRA combines 4-bit quantization, making it possible to fine-tune 70B parameter models on consumer-grade GPUs.58% similarUnverifiedNVIDIA designed a dedicated Tensor Core instruction mma.kind::mxf4 for NVFP4 in the Blackwell architecture (SM 100/103, such as the B100 and B200 series), and adopted an FP8 scaling factor mechanism with one scale per group of 16 elements.57% similarUnverifiedNVIDIA's Blackwell architecture (B100/B200) natively supports matrix operations in MXFP4 format at the hardware level, with throughput nearly doubling compared to FP8.57% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/697310API
curl https://kongchang.com/api/v1/knowledge/claims/697310MCP
get_claim(id=697310)