llama.cpp 0.7.0 Moves the Default Server Port to 9931

A cream llama silhouette connects through a blue route to a blue circular endpoint. A gray dashed route marks the previous endpoint, and a curved arrow indicates the move. The blue MacDigest binocular mark sits at upper left.

Conceptual illustration of llama-server’s default software endpoint changing. Illustration: MacDigest.

llama.cpp 0.7.0, released October 15, changes llama-server’s default listening port from 8080 to 9931. For Mac users running local models through a desktop client or development tool, the immediate upgrade check is whether that client and the server still use the same address.

Check your local client URL

The version-pinned server guide lists 127.0.0.1:9931 as the native default. For a client on the same Mac that accepts an OpenAI-compatible base URL, the corresponding address is http://127.0.0.1:9931/v1. The built-in browser interface uses the server address without /v1.

In practice, a saved client URL pointing at port 8080 will continue targeting that port. The server upgrade does not rewrite another app’s settings. Check saved connections, scripts and launch configurations before restarting, then confirm that a normal request reaches the intended server.

Keeping the previous endpoint is also an option: add --port 8080 to the existing server launch command, or set LLAMA_ARG_PORT=8080 in its environment. Command-line arguments take precedence over environment variables. An explicitly configured port therefore does not have to change.

Docker retains its container port

The port-change pull request explicitly preserves 8080 inside the project’s Docker server images. The tagged CPU Docker recipe, for example, sets LLAMA_ARG_PORT=8080. A Mac client connecting through Docker must still use the host port selected by its port mapping. Check that mapping before replacing an existing 8080 address with 9931.

Cache cleanup and Mac downloads

A separate prompt-cache fix stops embedding and reranking requests from filling the RAM cache with states that cannot be reused. Completion caching remains unchanged. The reported tests used CUDA hardware, so they do not establish a Mac memory-saving figure.

The formal 0.7.0 release links to nightly build b11669 for downloads. Both tags resolve to the same source commit, and that nightly includes separate macOS arm64 and x64 archives. Its download page is still marked pre-release. These are the same-commit nightly assets linked by the formal release.

MacDigest has not tested this release. Check the installed build and client configuration rather than assuming an updater has already delivered 0.7.0.

Leave a Reply

Your email address will not be published. Required fields are marked *