atomic.chat通过为LLaMA.cpp引入多令牌预测技术，大幅提升了本地大型语言模型的推理效率。该技术利用小型辅助模型预先生成后续令牌草案，由主模型进行验证。在MacBook Pro M5 Max上测试时，使Gemma 4 26B模型的令牌生成速度加快约40%，整体运行速度提升1.5倍。这项优化进一步巩固了LLaMA.cpp和GGUF格式在本地AI生态中的核心地位，为桌面应用、编程助手和私有设备助手等场景提供了更高效的部署方案。
在笔记本上把 Gemma 26B 的生成速度拉高 40% 是个真实的体验提升，atomic.chat 把 MTP 带入 LLaMA.cpp 生态，本地 AI 玩家可以直接拿去用。
atomic[.]chat just made Gemma 4 26B faster inside LLaMA.cpp.
making token generation about 40% faster in its MacBook Pro M5 Max test.
Great news for local llms, because LLaMA.cpp and GGUF sit close to the local AI user base, where support often spreads into desktop apps, coding agents, and private on-device assistants.
MTP (maltai token prediction) is like a smaller assistant drafting the next few words, while the main model checks whether those words are acceptable.
If the draft is correct, the system accepts several tokens quickly.
If the draft is wrong, the system rejects the wrong part and falls back to normal generation.
Multi-Token Prediction (MTP) for LLaMA.cpp! Running Gemma4 local model 1.5x faster. We patched LLaMA.cpp. Quantized Gemma 4 assistant models into GGUF format. We ran tests on a MacBook Pro M5Max. Gemma 26B with MTP drafts tokens 40% faster. Benchmarks, source code and models 👇在 X 查看被引用的帖子
来源：Rohan Paul · x.com