Can I run Qwen3-VL 8B?

Qwen3-VL 8B is Alibaba's vision-capable model in the 8B class, which makes it the one to reach for when you want a local assistant that can actually look at images and screenshots, not just read text. At a 4-bit quant it lands around 4.9 GB, so it fits on a 12 GB card like an RTX 3060 with room to spare and sits comfortably in unified memory on any Apple Silicon Mac. The minimum sensible footprint is about 8 GB of RAM, and if you are tight you can drop to a 2-bit build at roughly 3.4 GB, though the heavier q8 weighs in near 8.5 GB.

In daily use it feels quick for an 8B model. On an RTX 3060 you can expect around 63 tokens per second at 4-bit, and an M-series Max pushes that to roughly 72, both fast enough that replies stream faster than you read; a 4090 runs away at about 177. The headline number is the 256K context window, but treat it as a ceiling, not a default. At 128K context the model plus KV cache already needs around 21.7 GB total, well past what a single mid-range card holds, so keep working context modest unless you have a big GPU.

Against Llama 3.1 8B, which is the obvious same-size point of comparison, Qwen3-VL's standout is the thing Llama simply cannot do: it reads images. For plain text chat the two are close enough that the choice comes down to whether you need vision. If you only ever feed it text, Llama still has the deeper tooling ecosystem; if you want to describe screenshots or pull text out of a photo locally, this is the better tool. It ships under Apache 2.0, so you can use it commercially and in production without license worries.