llama.cpp 0.7.0: GPU cache for offloaded MoE experts, default port now 9931

llama.cpp v0.7.0 was published on GitHub on October 15, 2026. It matches nightly build b11669 and updates ggml to v0.27.0. The main addition is `--moe-cache-mib`: MoE experts kept in host memory run on the GPU from an LRU cache. The notes claim up to ~2.2x faster decode when experts are offloaded (#29887). In that PR, the 2.2x run is an RTX 5090 with all experts in host memory; RTX 4090 runs show 1.57x to 1.62x, all on one Qwen3.8-Flash-Next Q4_0 model. The PR suggests a cache of about 10% of total expert size. The cache also works across multiple GPUs (#30112). llama-server now listens on port 9931 by default instead of 8080. Docker images keep 8080 inside the container (#30159). Internally, chat formatting now produces an array of parts instead of a single string (#30389), and the chat layer is refactored around `common_chat_session` (#30210). Embedding models skip the logits buffer and the LM head when pooling is enabled. The notes say this uses up to 10x less VRAM (#30255, #30445). A new `--spec-draft-n-ubatch` flag sizes the draft context batch separately to save VRAM (#30459). LiliCorr and GLM5-Next MTP speculative decoding are added. New models include K2 Horizon, Prism Bonsai 2 27B (ternary), EmbeddingGemma-2, MiniCPM-V 4.7, Nemotron Super 3.5 VL, Nemotron Bolt 74B A9B and five decision models. The web UI gets a Models Manager with load/unload, delete from disk and resumable downloads (#29583). Later nightly builds added Muse Spark Open (b11671, Oct 16), Aleph Alpha's Kolibri-1, a MoE with 78B parameters and 3.46B active (b11709, Oct 17), Hugging Face Hub model discovery and download in the web UI (b11714, Oct 17), and `--connect` for peer-to-peer access to the server from outside the LAN (b11744, Oct 19). Its help text marks it experimental and off by default.