Qwen3.6-35B-A3B tops local agent benchmarks

If these results hold up, Qwen3.6-35B-A3B is more interesting as an agent runtime than as a pure benchmark champ: speed plus broad framework compatibility is exactly what local builders need.

• –Hermes looks like the real stress test here; passing the hardest harness matters more than scoring well on forgiving frameworks.
• –smolagents appears to flatter weaker models because it leans on code generation instead of strict structured tool calling, so its 100% numbers are not directly comparable.
• –The low HumanEval/MMLU numbers on Qwen3.6 likely reflect day-0 quantization or evaluation noise more than the model’s true ceiling.
• –Qwen3.5-35B at 8-bit reads like the safer all-rounder if you can spare RAM; Qwen3.6 is the speed pick.
• –The broader takeaway is that local agent stacks are now mostly a framework compatibility problem, not just a model-quality problem.