Qwen3.5-122B-A10B thinking mode trails GPT-OSS

This smells like a reasoning-budget problem, not a bandwidth problem: Qwen3.5’s default thinking pass can dominate wall time on easy extraction jobs.

• –Qwen3.5-122B-A10B is a 122B-parameter MoE with 10B activated, so raw tok/s can look fine even when extra reasoning tokens inflate end-to-end latency.
• –The model card explicitly says thinking is on by default and documents a non-thinking configuration, which is the lever you want for low-latency pipelines.
• –For tagging and other narrow extraction tasks, benchmark thinking and non-thinking modes separately; otherwise you’re measuring reasoning budget, not just model quality.
• –If latency is the blocker, GPT-OSS-120B is probably the safer fast-path default, while Qwen3.5 stays the better choice for harder prompts.