Version 2.20.6
2026-08-28Tool calling (function calling) support and a new guide for agent benchmarks.
New
- Models can use tool calling (function calling), vision models included.
- A guide for creating agent and tool-calling benchmarks (Turkish and English), with a working example.
Improvements
- When a model starts, its context length is set from the free GPU memory, so long-context models fail to start less often.
- A model that fails to start is stopped automatically.