Qwen 3.6 27B coding performance fails real-world tests

Qwen 3.6’s supposed dominance in coding benchmarks is meeting reality, and the result is a significant disconnect between synthetic evals and agentic utility.

• –Model fails to use standard filesystem tools, opting instead for unreliable Python-based text replacement scripts that lead to file corruption.
• –High SWE-bench scores (77.2) do not appear to translate into stable repository-level reasoning in real-world local environments.
• –Circular reasoning loops suggest that "Thinking Preservation" features may need further tuning for multi-turn developer workflows.
• –Local inference via oMLX on Apple Silicon remains a niche but revealing testbed for frontier model stability.