Mitii AI Coding Agent Hits 78% Success Rate on 500+ Tasks Using Local Qwen3-Coder 30B

Mitii, an open-source AI coding assistant built for fully local execution, was manually benchmarked across 515 tasks using the Qwen3-Coder 30B model running via Ollama, with no data sent to external APIs. The agent passed 400 of 515 tasks, achieving an overall 78% success rate across three operational modes: Agent, Plan, and Ask. Ask Mode performed strongest on hard tasks at 87%, largely due to its built-in impact analysis that requires user confirmation before modifying code. Security-related tasks saw an 87% pass rate, while the weakest area was medium-difficulty semantic retrieval, which scored 63%. The benchmark highlights that locally run large language models can be viable for production-grade coding agents, though improvements in context routing remain a priority.