Unverified50% confidenceBenchmarkExact time
在27B混合架构、4B稠密模型、36B MoE三类模型上以8k上下文测试，Prefill阶段性能无变化，Decode阶段打补丁后快约1.4%
1
Sources
50%
Confidence
Long-term
Relevance
7/25/2026
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 MiniCPM series achieves performance close to 7B and even 13B models on multiple benchmarks with a parameter count at the 2B–4B level71% similarUnverifiedThe 14B MoE model is derived by compressing the Qwen3.6 35B MoE, reducing total parameters from 35B to 14B, while the active parameters per forward pass remain at 3B, fully consistent with the original 35B69% similarVerifiedThe larger the number of model parameters, the greater the robustness to quantization, allowing 70B-level models to maintain performance close to FP16 after 4-bit quantization.69% similarUnverified混元3是混合专家模型（MoE），总参数量为295B，激活参数约21B68% similarUnverified千问3.6的35B MOE模型每次只激活其中3B的参数68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/613152API
curl https://kongchang.com/api/v1/knowledge/claims/613152MCP
get_claim(id=613152)