llama.cpp v0.7.0 adds LiliCorr, MoE GPU cache, many new models
Release v0.7.0 adds LiliCorr and GLM5-Next MTP speculative decoding, a host-memory GPU MoE cache (--moe-cache-mib, up to 2.2x decode), many new models, server API changes and ggml v0.27.0.
What changed
Added LiliCorr speculative decoding and MTP for GLM5-Next; new GPU host-memory MoE cache (--moe-cache-mib, up to 2.2x decode); refactored common_chat_session API; server default port 9931; web UI models manager; ggml v0.27.0.
Why it matters
Developers can reduce VRAM use for embedding workloads up to 10x.
Who should care
DevelopersResearchersEnterprises
Worth trying
Yes — Try --moe-cache-mib for MoE offload; new models and server APIs are available.