Unverified50% confidenceBenchmarkExact time
Using the Qwen3-27B model on the wikitext-2 dataset with fp32 reference logits as a baseline for KL divergence measurement, the median KLD dropped from 0.0023 to 0.000001, an approximately 2300-fold improvement in precision.
1
Sources
50%
Confidence
Long-term
Relevance
7/24/2026
First Seen
Sources
三行代码修复：Tesla P100在llama.cpp中精度提升2300倍
redditr/LocalLLaMA7/12/2026
Related Claims
UnverifiedThe Qwen3.6 34B 80-layer model achieves an MTP speculative decoding acceptance rate of approximately 78% on LM Studio, while the original 27B achieves only approximately 57–58%67% similarUnverifiedGlimmer's intelligence index score of 35 surpasses Gemma 2 27B (30) and nearly matches Qwen 2.5 (36), with Qwen 2.5 72B topping the leaderboard at 38.66% similarUnverifiedK-Quant是llama.cpp生态中的混合精度量化方案，根据每层权重敏感度分配不同量化bit数，可在相同平均bit数下将精度损失降低40-60%66% similarUnverifiedQwen3.6-27B's benchmark scores place it at the leading position among models of similar scale (20B–30B parameter range).66% similarVerifiedQwen3-235B has a total parameter count of 235 billion, adopts a MoE architecture, and activates approximately 22B parameters during each forward inference pass.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/608968API
curl https://kongchang.com/api/v1/knowledge/claims/608968MCP
get_claim(id=608968)