User finds Qwen3.5 122B superior for complex tool calling tasks

A Reddit user reports that the Qwen3.5 122B model outperforms smaller Qwen and Gemma variants in handling complex, multi-step tool calling workflows.

• The user tested Qwen3.6 27B and 33B models, finding they failed on complex tasks by being overly aggressive and digging themselves into problems.
• Gemma4 31B and 26B models were described as unable to complete tasks independently, requiring constant manual intervention.
• Tool calling functionality in both Qwen and Gemma MoE models was reported as buggy, with consistent blank responses.
• Qwen3.5 122B successfully extracted specific data fields from approximately 160 PowerPoint presentations in about two hours without issues.
• The user concludes that ~120B size MoE models offer the best balance of capability and speed for local deployment on systems with sufficient RAM.