llama.cpp can convert MiMo-V2.6 Pro and Flash one day after release.

PR #29257 moved the Kimi-K3 mxfp4 expert repack into base.py so DeepSeek-V4, Kimi-K3 and MiMo share it. Both models load with vision and needed no runtime changes; the mmproj audio decoder is excluded. Separately, --host now takes a comma-separated list of addresses and Unix sockets, so you can serve localhost and a Tailscale address at once without a reverse proxy, and llama-server's router stopped evicting models mid-load and proxying requests into dying child processes.