Qwen3.5-35B-A3B hits 13.2 t/s on 8GB

This is the kind of post that matters more than another benchmark chart: it turns a “too big for laptop” model into something actually usable, but only by leaning hard on quantization, offload tuning, and conservative context settings.

• –The official model card says Qwen3.5-35B-A3B has 35B total parameters, 3B activated, and 262,144 native context, so the model is explicitly built for efficiency rather than brute-force density.
• –The reported speed depends on a narrow setup: IQ3_XXS quantization, `-ngl 18`, 6 threads, Vulkan, and an 8K context. Any of those knobs moving the wrong way will likely drag throughput down fast.
• –The Unsloth imatrix quant is the real unlock here; this is less “35B on 8GB in general” and more “the right quant makes a very specific local deployment possible.”
• –For local-LLM tinkerers, the takeaway is that backend and quant choice matter as much as model choice. On laptops, inference engineering is the product.
• –Community replies already point at the obvious next experiments: let auto-fit handle more of the layer placement, lower context, or accept that smaller Qwen3.5 variants will be much easier to scale.