Can I run Qwen 3.6 35B-A3B?

Qwen 3.6 35B-A3B is a mixture-of-experts model, and that is the whole story of who it is for. It carries 35B total parameters but only activates 3B per token, so it runs at the speed of a tiny model while needing the memory of a big one. At a 4-bit quant the weights land around 21.2 GB, but the model still needs roughly 32 GB of RAM minimum to hold the full set of experts. That rules out a 12 GB card like the RTX 3060, where it does not fit. The natural homes are a 24 GB RTX 4090 or an Apple Silicon Mac with plenty of unified memory.

In daily use the active-3B math pays off: on a 4090 you can see around 471 tokens per second, and on an M-series Max around 192, which is faster than most dense models a third of this size. Even CPU inference on DDR5 manages about 28 tokens per second, usable for batch work. It handles chat, reasoning, coding, and vision, so it earns its keep as a single do-everything local model. The 256K context window is generous, but memory is the catch: filling just 128K already pushes total footprint to about 53.8 GB, so long-context work wants a 64 GB machine, not a 24 GB card.

Against the dense Command R 35B in its related lineup, this MoE generally feels much faster at the same nominal size, since you pay for 35B of memory but compute like a 3B. The tradeoff is the usual MoE one: a dense 35B can sometimes hold an edge on the hardest single-shot reasoning, though for most work the speed wins. Its standout trait is that speed-to-capability ratio plus native vision in one package. License is Apache 2.0, so you are free to use it commercially and in production with no provider strings attached.