wrldsuksgo2mars/DeepSeek-V4-Flash-Vision-Exp-EXL3-K2.2-D2-v1 with the refusal edit from apetersson/DeepSeek-V4-Flash-Vision-Exp-Abliterated applied: the 33 layers.10..42.attn.wo_b tensors were replaced with the abliterated ones.
42 tok/s on code, 30 tok/s on prose, at a 500k context window with dSpark on a GB10.
Credits: DeepSeek, wrldsuksgo2mars, apetersson, drowzeys. MIT.
model-00007-of-00011.safetensors 8.6 GBmodel-00010-of-00011.safetensors 8.6 GBmodel-00003-of-00011.safetensors 8.6 GBmodel-00009-of-00011.safetensors 8.6 GBmodel-00005-of-00011.safetensors 8.6 GBmodel-00004-of-00011.safetensors 8.6 GBmodel-00006-of-00011.safetensors 8.6 GBmodel-00008-of-00011.safetensors 8.6 GBmodel-00001-of-00011.safetensors 8.6 GBmodel-00002-of-00011.safetensors 8.6 GBmodel-00011-of-00011.safetensors 5.9 GB有文件?提交种子。
提交种子