llama.cpp adds MTP context KV cache support for DeepSeek-V2 and GLM-4-MoE
September 11, 2026
The llama.cpp repository merged a pull request to fix MTP context KV cache allocation. This update enables optimized inference for DeepSeek-V2 and GLM-4-MoE architectures within the ggml runtime.
HOW THIS AFFECTS YOU
●
builderYou can now run DeepSeek-V2 and GLM-4-MoE with improved context management in llama.cpp.