Bonsai 27B field guide
Ternary Bonsai 27B passed a structured tool call
The exact local tool call check, what passed, and what still needs testing before an agent can take actions.
Does Ternary Bonsai 27B support structured tool calls?
Yes in our basic local check. Through LM Studio's OpenAI compatible endpoint, the model selected get_weather, supplied Lisbon as the city, and ended with the tool_calls finish reason. This proves the response shape worked for one call. It does not prove a long agent loop is reliable.
What passed
The request included one function with one required string field. The model returned a structured function name and valid JSON arguments instead of writing a fake call in plain text.
What to test next
- Return tool output to the model and confirm it writes the final answer.
- Test several tools with similar names.
- Test missing arguments and tool errors.
- Measure success across many full loops, not one request.
Questions people ask
Does one passing call make it ready for an autonomous agent?
No. An agent also needs planning, error recovery, and stable behavior across many tool turns.
Did the call use a cloud API?
No. The model ran locally. The test tool was a schema check and did not fetch live weather.
Primary sources
Vendor claims on this page are labeled as PrismML claims. Local results are added only after a saved trace completes.