MoE models need different prompting for agentic workloads. The routing layer decides which experts activate per token, and tool-call JSON can land on suboptimal expert paths if your system prompt is not structured right. Try explicit XML-style tool schemas instead of free-form JSON. Qwen3 dense models avoid this because every param sees every token. Not a model quality issue, it is a routing architecture issue.

MOE routing is the bottleneck for agentic tasks. The model needs to pick the right expert on every token, and tool-use prompts are out of distribution for most training mixes. Total params matter less than how well the router was trained on structured output.