Qwen3.5 122B Outshines Rivals in Complex Tool Calling Tasks
After testing five models, one 122B MoE model completed 160 PowerPoint extractions in 2 hours.
A developer with 128GB RAM exhaustively tested five AI models for a complex data extraction task: pulling specific fields from 160 PowerPoint files using tool calls. The smaller Qwen3.6 27B and 33B models turned aggressive, performing slightly more than asked and digging themselves into problems. Gemma4 31B and 26B were the opposite—too passive, requiring constant user supervision ("ok, ok, ok"). Both Qwen and Gemma MoE models also exhibited buggy tool calling with blank responses.
By contrast, Qwen3.5 122B completed the entire task autonomously in about two hours after a single goal prompt. The author argues that fully dense ~30B models are too slow, while MoE models around that size are unreliable. The ~120B MoE size hits a sweet spot of capability and speed, recommending users with enough RAM to skip smaller models and either use APIs or this larger local size. The post highlights a growing sentiment that for complex multi-step tasks, model scale matters more than architecture tricks.
- Qwen3.5 122B succeeded in extracting data from 160 PowerPoints in 2 hours autonomously, while smaller models failed or required constant supervision.
- Qwen3.6 27B/33B models were aggressive and self-destructed on complex multi-tool tasks; Gemma4 31B/26B were too passive and needed babysitting.
- The author recommends ~120B MoE models as the sweet spot for capability and speed, noting smaller dense models are too slow and smaller MoE models are unreliable.
Why It Matters
For professionals needing reliable AI for complex multi-step tasks, 122B MoE models offer autonomous efficiency over smaller alternatives.