XiaomiMiMo

/

MiMo-VL-7B-RL

like 108

Follow

Xiaomi MiMo 603

thanks for your excellent MiMo-VL-7B-RL model.

I build a MiMo-VL-7B-RL-q4_k_m.gguf model in https://huggingface.co/zhouwg/kantv/tree/main by the official tools provided in llama.cpp.

then I compared it with Qwen1.5-1.8B,Qwen2.5-3B,Qwen3-4B,Qwen3-8B,gemma-3-4b,gemma-3-12b,Llama-3.1-Nemotron-Nano-4B,Phi-4-mini-reasoning,DeepSeek-R1-0528-Qwen3-8B on my Qualcomm Snapdragon 8Elite based phone, I can't believe that it achieved the second best overall experience at the moment:

for some questions, the MiMo-VL-7B-RL is much better than DeepSeek-R1-0528-Qwen3-8B and Qwen series, Gemma-3-4B is a little better than MiMo-VL-7B-RL.
the MiMo-VL-7B-RL also gave some stupid answers to some simple questions, DeepSeek-R1-0528-Qwen3-8B is better on these questions, Google's Gemma-3-4B achieved the best overall experience.

my question is:

as an Image-Text-to-Text multimodal model, where is the mmproj model file? should we create the mmproj model file manually? how to create the mmproj model file?

thanks.