llama.cpp has tagged build b8808, and while it's not a headline-grabbing model drop, it addresses a real server-side rough edge — the media marker used internally by the server is now randomized rather than static, closing off a class of potential conflicts with model output.
What's new
The core change in b8808 is in the server component: media markers are now generated randomly instead of relying on a fixed string. Alongside that, the legacy <__image__> token has been removed entirely. The PR also includes minor nits and a revert on a special character that snuck into the random marker implementation. Small diff, focused scope.
Why it matters
Static media markers are a collision hazard — if a model happens to produce the same string the server uses internally as a delimiter, things break in ways that are annoying to debug. Randomizing the marker is the sensible fix. Dropping the legacy image token cleans up technical debt that's been sitting in the codebase for a while.
What to watch
Builds are available for the usual targets: macOS Apple Silicon (with and without KleidiAI), macOS Intel, iOS XCFramework, and Linux on x64, arm64, and s390x. If you're running a local multimodal server setup, this is worth pulling sooner rather than later.