Ternary Bonsai 27B passed a structured tool call

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.

Sources

• PrismML Hermes integration guide
• PrismML local server guide
• Ternary Bonsai 27B MLX model card

Related Bonsai 27B lessons

• How to use Ternary Bonsai 27B with Hermes Agent
• Ternary Bonsai 27B needs room for reasoning tokens
• Ternary Bonsai 27B limitations and verification checklist

The benchmark numbers on this page describe one checkpoint, runtime, machine, and test shape. Reproduce the test on the hardware and workload you plan to use before making a product decision.