Abiray

/

Minimax-H3-nvfp4-INT4-INT8-Convrot

like 95

Hi, good news! The I2V model is actually already uploaded. MiniMax officially names their Image-to-Video model FL2VA (First/Last Frame to Video/Audio).
You can find both the pure INT8 version and optimized 80/20 mixed-precision version (INT4 + INT8) already in the repo:
MiniMax_H3_FL2VA_pruned_mixed_int4_int8_convrot.safetensors
MiniMax_H3_FL2VA_pruned_int8_convrot.safetensors

Hi, good news! The I2V model is actually already uploaded. MiniMax officially names their Image-to-Video model FL2VA (First/Last Frame to Video/Audio).
You can find both the pure INT8 version and optimized 80/20 mixed-precision version (INT4 + INT8) already in the repo:
MiniMax_H3_FL2VA_pruned_mixed_int4_int8_convrot.safetensors
MiniMax_H3_FL2VA_pruned_int8_convrot.safetensors

you're the best, thank you SO much !!!

Any chance the I2V model, pruned, is going to be pushed as nvfp4?

Sure

can u add details on what GPu can be used with which model for noobs like me.

done

Where are the models, why were they deleted? MiniMax_H3_FL2VA_pruned_mixed_int4_int8_convrot.safetensors
MiniMax_H3_FL2VA_pruned_int8_convrot.safetensors 😥😲

I had to delete MiniMax_H3_FL2VA_pruned_mixed_int4_int8_convrot.safetensors because it had some issues, but I'll be updating it soon. The MiniMax_H3_FL2VA_pruned_int8_convrot.safetensors file is still there.

2. Text Encoders (/text_encoders)
   qwen3vl_32b_minimax_h3_int4_convrot.safetensors (15.0 GB) - Recommended for 16GB GPUs
   qwen3vl_32b_minimax_h3_int8_convrot.safetensors (27.1 GB) - Recommended for 24GB GPUs
   qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (15.7 GB) - For Blackwell GPUs

I would've thought the Blackwell GPUs (32GB-96GB VRAM) would be recommended for the largest file. Should the last two recommendations or precisions be swapped?

I had to delete MiniMax_H3_FL2VA_pruned_mixed_int4_int8_convrot.safetensors because it had some issues, but I'll be updating it soon. The MiniMax_H3_FL2VA_pruned_int8_convrot.safetensors file is still there.

MiniMax_H3_FL2VA_pruned_int4_convrot.safetensors produces garbled results, i think there is something wrong with it aswell. The replaced MiniMax_H3_FL2VA_pruned_mixed_int4_int8_convrot.safetensors works fine though

qwen3vl_32b_minimax_h3_int4_convrot.safetensors这个文件有报错

qwen3vl_32b_minimax_h3_int4_convrot.safetensors这个文件有报错

use this https://huggingface.co/Abiray/MiniMax-H3-GGUF/blob/main/text_encoders/qwen3vl_32b_minimax_h3-Q4_K_M.gguf

FL2VA nvfp4 please~

FL2VA nvfp4 please~
done