I Tested 6 Local Models on Real Agent Tasks. The Best Scored 50%.

A benchmark of six local models found that strong coding performance did not translate into agent-task success.

SmolLM3-3B led the test at 50%, while Phi-4-mini, Qwen2.5-Coder-14B, and Llama 3.1-8B fell well below their code-benchmark scores.

The custom benchmark checked tool selection, chaining, argument passing, and whether models stayed silent when no tools were available.

A 100-line proxy converted text tool outputs into OpenAI-style tool calls for evaluation, highlighting the gap between coding skill and agent readiness.