A better local AI model by default
New installs that run AI on their own server now use Qwen 3 4B Instruct instead of Llama 3.2 3B. On an ordinary 4-core CPU it handled twice as many agent jobs correctly (12 of 15 against 6) at about five seconds a reply. Existing installs keep the model they have; switch in Admin, AI.
- Dates the AI gets right. Agents and the assistant now see the coming days written out, so "next Friday" lands on Friday. Small models counted wrong from today's date alone.
- No more refused AI requests with reasoning on. OneCamp now asks the model engine whether a model can reason before turning reasoning on, instead of guessing from its name. Turning reasoning on with a model that cannot do it used to make every AI request fail on newer Ollama.
v2.41.0 with AI