Unverified90% confidenceFactTime unknown
Meta的LLaMA官方实现采用非交错格式（half-half格式），GPT-NeoX/EleutherAI的实现采用交错格式
1
Sources
90%
Confidence
Long-term
Relevance
6/1/2026
First Seen
Sources
DeepSeek-V3.2-Exp发现RoPE实现Bug：交错格式不匹配导致推理性能下降
twitterdeepseek_ai
Related Entities
Related Claims
UnverifiedMeta在2023年2月发布LLaMA，首次将接近GPT-3水平的大语言模型权重公开70% similarUnverifiedOpen-source models such as Meta LLaMA 3, Mistral, and Google Gemma have capabilities approaching those of early GPT-4, and can run smoothly on an RTX 4090 for models up to 70B parameters when paired with inference frameworks such as Ollama and llama.cpp.67% similarUnverifiedLLaMA是Meta开源的大语言模型系列，以较小的参数量实现了接近GPT-3.5的性能65% similarUnverifiedLLaMA是Meta开源的大语言模型系列，以较小的参数量实现了接近GPT-3.5的性能65% similarUnverifiedLLaMA是Meta开源的大语言模型系列，以较小的参数量实现了接近GPT-3.5的性能65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/27879API
curl https://kongchang.com/api/v1/knowledge/claims/27879MCP
get_claim(id=27879)