Qwen3.5-35B-A3B courts Opus switchers

Hot take: this is less about raw hardware and more about workflow tolerance. On a machine that big, Qwen3.5-35B-A3B is plausibly a strong local default, but Opus still has the safer margin when tasks get messy.

• –The model is built for efficiency: 35B total parameters, 3B activated, Apache 2.0, and native 262k context make it attractive for local agent use.
• –Hugging Face’s published results show it is competitive on coding and agent benchmarks, but not an across-the-board win over frontier closed models.
• –For daily coding, it likely covers the common path well: edits, refactors, test writing, and tool use; the gap shows up in deep debugging, ambiguous specs, and multi-step reasoning.
• –On an M5 Max 128GB, memory is probably not the limiting factor; latency, quant quality, and context management will decide whether it feels “good enough.”
• –Best answer for most teams is probably a split setup: Qwen for routine local work, Opus for the hard cases.