Unverified50% confidenceFactExact time
In the Qwen3.8 27B vLLM tests, using the Marlin kernel to keep quantized weights while keeping activations in bf16 lowered NVIDIA NVFP4's KLD from 0.0256 to 0.0117, a reduction of more than half.
1
Sources
50%
Confidence
Long-term
Relevance
11/7/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn tests on Qwen3.8 27B, the 8-bit KV cache of llama.cpp and ExLlamaV3 (q8_0 and EXL3 8-bit) produced KL divergence changes within ±0.00012 across all tested weight quantization versions, falling within the range of numerical noise.71% similarUnverifiedThe NVFP4 version of Qwen3-27B shows only marginal loss compared to uncompressed FP16 on the MMLU Pro and GPQA Diamond benchmarks.70% similarUnverifiedOn the RTX 4070 Ti SUPER, NVFP4 runs via the weight-only Marlin kernel, with weights kept in low-precision storage but the actual computation still performed at 16 bits.69% similarUnverifiedThe Q4_K_M/GGUF version of Qwen 27B decodes at approximately 10.7 tok/s on the same machine, roughly 1/3.8 of the MLX version.68% similarUnverifiedThe 3-bit and NVFP4 versions of Qwen3.8-27B scored 0.39 and 0.34 respectively, trailing the MoE-architecture Flash-Next, with a paired difference of +0.16, P = 0.92.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1146820API
curl https://kongchang.com/api/v1/knowledge/claims/1146820MCP
get_claim(id=1146820)