Can I run Qwen3-VL 30B-A3B?

Qwen3-VL 30B-A3B is Alibaba's mixture-of-experts vision model, and the headline trick is in the name: 30B total parameters but only 3B active per token. That means it runs at the speed of a tiny model while still needing memory for the whole thing. At a 4-bit quant it lands around 18.2 GB, and you want at least 32 GB of RAM to hold it comfortably. It does not fit on a 12 GB card like an RTX 3060, but an Apple Silicon machine with plenty of unified memory or a 24 GB GPU is its natural home. It is built for people who want image understanding plus chat and reasoning locally.

In daily use the active-3B design pays off: on an RTX 4090 you can see around 471 tokens per second, and even an Apple M Max stays brisk at roughly 192 tok/s, fast enough that vision answers feel instant. On a CPU with DDR5 it drops to about 28 tok/s, usable but no longer snappy. The 256K context window is the marketing ceiling, not a free lunch. Filling it is expensive: at 128K context the model plus cache climbs to about 48.6 GB total, so plan your memory around the context you actually use rather than the maximum on the spec sheet.

Against Gemma 4 31B, the closest related model here, the two sit in similar territory for size, but Gemma 4 is a dense 30.7B that activates every parameter, so Qwen3-VL generally feels faster for its footprint while Gemma tends to be steadier on pure reasoning and coding. Qwen3-VL's standout trait is that vision capability bundled into a model this quick on consumer hardware, which is still uncommon locally. It carries an Apache 2.0 license, so you can use it commercially and in production without provider-specific restrictions. Pull it with the qwen3-vl:30b Ollama tag and go.