Qwen3.5-4B Counts Better With Thinking

The result is a useful prompting lesson, not evidence of universal counting competence.

• –Thinking mode turns a brittle visual or character-counting task into a more reliable sequential process
• –At 4B parameters, Qwen3.5-4B offers an unusually capable local, open-weight option for experimentation
• –The improvement comes with added latency and token usage, making thinking best reserved for error-prone tasks
• –Five successful trials are encouraging but too small to qualify as a robust benchmark
• –Developers should pair model reasoning with deterministic code when exact counting matters