VOLlama v1.0.0-beta.1

Changelog

VOLlama 1.0 rebuilds much of the architecture and adds some big new features, so the beta needs thorough testing.

It is important to run Chat > Reset and start fresh.

Streamlined

Apple is winding down support for Intel Macs, so VOLlama's release binaries will support only Macs with Apple Silicon. To run VOLlama on an Intel Mac, run it from source or build it on an Intel machine.

The API Settings dialog is gone. VOLlama now only uses OpenAI compatible chat completion API, which Anthropic, Gemini and OpenRouter all support. Embedding goes through an OpenAI compatible API as well.

VOLlama now depends far less on LlamaIndex. There is mno more platform specific library, and this means fewer update is necessary when there's new model. Also it does not rely on the LlamaIndex web readers.

Presets now hold the wider configuration: base URL, API key, model, context window, system prompt, generation parameters and the RAG settings.

Preset Manager (Chat > Presets Ctrl/Cmd+P) lets you create, edit, duplicate and delete presets.

Agentic tools

This one is dangerous and experimental, but exciting feature. You can ask the model to write a script, look through files, analyze data, and so on.

Inspired by the pi agent, VOLlama gives the model a small set of tools to run commands and read, write and edit files, if you turn on Chat > Tools.

It does not ask for approval before each command, and there is no undo. It asks you to agree every time you turn them on. Only turn it on for a model you trust — small local models call tools badly.

In Chat > Workspace, you can pick the folder the tools work in. The default is VOLlama in your home folder. It is a starting point, not a sandbox, so the model can run a command elsewhere and write outside it.

Other improvements

• Reasoning: Chat > Show Reasoning prints the model's reasoning in the chat.
• Compaction: long chats are compacted automatically before the server runs out of room. You can manually run Edit > Compact Conversation (Ctrl/Cmd+Shift+K).
• Once an index is loaded or built for RAG, the model can search it with a tool call. You no longer have to prefix a question with /q, though /q still works for models withouth toolcall support.
• RAG > Show Context prints the passages a search retrieved.
• RAG > Clear Index unloads the index and disables the search tool.
• Voices are grouped into submenus by language rather than by identifier.
• When the server reports a cached prompt, speed and token counts are calculated from what it actually processed, so a cache hit no longer impacts the speed calculation.