Can I run GLM-4.6V-Flash?

GLM-4.6V-Flash is a 9B vision-and-chat model from Z.ai, aimed at people who want a local assistant that can actually look at images, not just read text. At a 4-bit quant it weighs about 5.5 GB, so it slots onto a 12 GB card like an RTX 3060 with room to spare and fits inside unified memory on any reasonably specced Apple Silicon Mac. The minimum you want is around 12 GB of RAM to load it comfortably. If you are squeezed for space, a 2-bit build drops to roughly 3.8 GB, though you trade away quality for that headroom.

In daily use it feels quick. On an RTX 3060 you can expect around 56 tokens per second at 4-bit, fast enough that replies stream past reading speed, and a 4090 pushes that to roughly 157 tok/s if you have one. An M-series Max sits near 64 tok/s, while pure CPU on DDR5 crawls at about 9 tok/s, usable only if you are patient. The 128K context is genuine but expensive: filling it pushes total memory toward 23.2 GB, well past what a 12 GB card holds, so keep working context modest and lean on vision tasks rather than enormous documents.

Against its siblings, Qwen 2.5 VL 7B is the obvious comparison point on the vision side and generally holds up well, while Qwen 3.5 9B tends to be the stronger pick if you also want reasoning alongside images. GLM-4.6V-Flash's standout trait is being a genuinely compact multimodal model that runs on modest hardware without much fuss. The license is MIT, which means you are free to use it commercially and in production with no provider strings attached, a real advantage over many open-weight releases.