Bumblebee 7B V1.5 is out!

[MonolithFoundation]

Bumbebee 7B V1.5 make use of Qwen2 7B model, and we got a surprising result compare with Bunny LLama3 8B and MiniCPM LLama3 8B:

Model	MMB-CN-Test	MMB-EN-Test	MMStar(A)	MMStar(C)	MMStar(F)
Bumblebee-14B	75.8	76.8	43.8	63.2	41.2
Bumblebee-7B-V1.0	70.5	71.8	40.4	66.2	33.6
Bumblebee-7B-V1.5	76.5	77.29	46	65.6	43.2
Bunny-Llama3-8b	73.9	77	43.5	-	-
MiniCPM-Llama3	73.8	77.6	-	-	-

In short, our model surpassed Bunny on both Chinese and English, but the most important part is, our model uses only 144 tokens per image, while Bunny uses 576 of them, and for MiniCPM, even more!

Leave a comment below if you wanna discuss about the detail!