Qwen3.6-35B-A3B Q4_K_M posts solid CPU evals

The headline result is that a 35B/3B MoE model stays quite usable after quantization on CPU, but the weaknesses show up exactly where you’d expect: precise code and tool-use tasks degrade faster than broad commonsense reasoning.

• –22 tokens/sec on a no-GPU host is the real story here; this is practical enough for local experimentation and some interactive workflows.
• –HellaSwag at 74.3% suggests the quant preserves general language competence better than task-specific exactness.
• –HumanEval at 47.56% and BFCL at 46% imply that quantization still hurts structured output and function-calling reliability.
• –The setup matters: this is a realistic local deployment path, not a cherry-picked GPU benchmark.
• –The eval is useful as a deployment signal for people deciding whether Qwen3.6-35B-A3B is viable on commodity CPU infrastructure.