Deepseek Vision lacks perspective and reference to the size within the same image

So, recently DeepSeek updated vision. It still doesn’t know proportion. For example, if you place an M.2 B-key 3042/3052 beside an M-key. It will not be able to pick up the size difference of 2 slots.

Any human will be able to tell the empty slot is a rare B-key m.2, but the DeepSeek will just say it’s a 2242 slot like most AI.

I know that a B-key isn’t a common M.2 slot. So training data may be limited. But there is literally an example of an M-key on top with a 2280 M.2 slotted in as a reference.

So basically all the B-key 3042/3052 get hidden or just buried under the noise of 22 mm M.2 slots.