Mistral Small 4 trips over agents

The model looks strong on paper, but agentic coding lives or dies on wrapper compatibility more than benchmark slides. If a model can’t survive tool round-trips cleanly, it won’t feel smart in daily use, no matter how good the weights are.

• –Mistral positions Small 4 as a 119B MoE model with 6B active parameters, Apache 2.0 licensing, and support for coding, reasoning, and multimodal workflows.
• –The user’s report is less about raw capability than about formatting failures, consecutive-message constraints, and tool-output parsing breaking common local agent stacks.
• –A commenter suggests LiteLLM as a translation layer between the agent CLI and the backend, which hints the serving/template stack may be the real bottleneck.
• –For local benchmarkers, this is a reminder to score “agent compatibility” separately from coding accuracy, because a brittle chat template can sink an otherwise capable model.