Qwen3.5 users weigh dense vs MoE

The 27B card makes the strongest practical case here: for coding, it is already close enough to the 122B MoE that latency is likely the bigger limiter than model size.

• –The official [Qwen3.5-27B](https://huggingface.co/Qwen/Qwen3.5-27B) card shows SWE-bench Verified at 72.4, basically matching [Qwen3.5-122B-A10B](https://huggingface.co/Qwen/Qwen3.5-122B-A10B) at 72.0 and beating it on IFEval at 95.0 vs 93.4.
• –The big MoE models buy ceiling, not free speed: [Qwen3.5-122B-A10B](https://huggingface.co/Qwen/Qwen3.5-122B-A10B) is 122B total / 10B activated, while [Qwen3.5-397B-A17B](https://huggingface.co/Qwen/Qwen3.5-397B-A17B) is 397B total / 17B activated.
• –For a workstation workflow, subjective speed matters more than benchmark bragging rights, so the 5090-style bandwidth upgrade sounds like the better quality-of-life move if 27B already feels close.
• –If you want a middle step, Qwen3.5-35B-A3B is the compromise model to test first, but I would still treat it as a throughput play rather than a reason to skip a fast dense setup.