mostlygeek/llama-swap is a tool for switching models on local OpenAI- and Anthropic-compatible servers. Release v269, published on November 1, 2026, adds a feature that stores captures on disk so they survive a restart. Embedding model support, PNG export of metrics, and sorting the model list by type are included as well. The source of these changes is the v269 release record on GitHub.
The scope of this project is switching models across local OpenAI- and Anthropic-compatible servers, including llama.cpp and vllm. The changes visible in v269 cover how captures are stored, which models are supported, the format for exporting metrics, and the ordering used to show models. So this release is better read as an update that adds these four features to an existing tool, rather than an introduction of a new switching tool. Captures are stored on disk and remain after a restart, metrics can be exported as PNG, and embedding model support and sorting by model type are both named in the release summary.
The provided material does not explain which settings or procedures each feature works through. The range of supported embedding models, the items included in the metrics, and the storage location and retention conditions for captures are not given either. No performance numbers or comparison against earlier versions can be confirmed. So the changes in v269 should be understood as limited to added features, and this material alone is not enough to judge their effect in real use or the range of compatibility.