3.50.0

AI Chat that works out of the box, and a refreshed local model list.

AI Chat loads its model for you (#1086)

Users downloaded Qwen or Gemma models and the chat stayed silent: nothing told them a model had to be loaded, and the Load button was buried in a settings dialog.

• Opening the AI Chat loads a model. It picks the last one you used, else a model recommended for the agent, else the first one downloaded. If you have none yet, it loads as soon as a download finishes.
• Nothing is loaded at app startup. A model can take gigabytes of memory, and most sessions never use the AI.
• Model picker in the chat header, like Ollama's. Switch models, Unload model (free memory), or Download models…. A status line explains what is happening: loading, a load error, or no model yet, with a one-click action.
• To turn auto-loading off, untick Load the last model when the AI Chat opens in AI → AI Model Settings.

Models now get their own chat format

Every model used to receive the same generic prompt format, whatever format it was trained on. On Qwen 3.5 that format never ended the model's turn: it answered, then kept inventing new user/assistant turns until it hit the token limit, so the agent got no usable reply. Prompts now use each model's own template (Qwen/ChatML, Gemma 2/3, Gemma 4, Llama 3, and anything else llama.cpp recognises), and the model answers once and stops.

The newer "thinking" models (Qwen 3.5/3.6, Gemma 4) are run with thinking switched off, so the agent's token budget goes to the answer.

Refreshed model list (all Apache-2.0)

Model	Download	Notes
Gemma 3 1B	0.8 GB	smallest option
Qwen 3.5 4B	2.7 GB	recommended agent model, 8 GB machines
Gemma 4 E4B	5.0 GB	all-rounder, 8 GB machines
Qwen 3.5 9B	5.7 GB	stronger agent model, 16 GB machines
Gemma 4 12B	7.1 GB	strong on long multi-step tasks, 16 GB
Qwen 3.6 27B	16.8 GB	top-tier, 24 GB
Gemma 4 26B-A4B	17 GB	mixture-of-experts, fast, 24 GB
Qwen 3.6 35B-A3B	22 GB	strongest, 32 GB

Gemma 3 4B/12B/27B, Llama 3.2 3B, Qwen 2.5 and Qwen3 2507 were removed from the download list. Copies you already downloaded still show up in the picker and keep working.