Qwen3.6-35B-A3B tops Mac Air picks

The bigger takeaway is that 32GB on a Mac Air is now enough for serious local LLM work, but only if you choose sparse models and a decent runtime instead of chasing dense parameter counts.

• –Qwen3.6-35B-A3B is the best "big-model feel" option here because its MoE design buys capability per GB that dense models at similar sizes usually cannot match.
• –Gemma 4 looks like the safer general-purpose fallback; Qwen3.6-35B-A3B is the sharper pick if the goal is agentic coding and tool-heavy workflows.
• –On Apple Silicon, backend choice matters almost as much as the model. MLX, llama.cpp, and Ollama can produce very different real-world speed and memory behavior.
• –The practical sweet spot for a 32GB MacBook Air is still the 27B-35B class; go much larger and latency plus KV-cache pressure start eating the value.