aleada/Pixtral-12B-W4A16
Image-Text-to-Text • 13B • Updated • 91 • 1
W4A16 for consumer Ampere
Weight-only INT4 for RTX 3090/4090-class cards, where FP8 and NVFP4 are emulated or unusable. Full recipe on every card.
-
-
aleada/Llama-3.2-11B-Vision-Instruct-W4A16
Image-Text-to-Text • 11B • Updated • 120 -
aleada/Gemma-3-12B-it-W4A16
Image-Text-to-Text • 12B • Updated • 38 -
aleada/Phi-4-W4A16
Text Generation • 15B • Updated • 350 -
aleada/Phi-4-mini-instruct-W4A16
Text Generation • 4B • Updated • 373 -
aleada/SmolLM3-3B-W4A16
Text Generation • 3B • Updated • 330 -
aleada/Gemma-4-12B-it-W4A16
Image-Text-to-Text • 12B • Updated • 117
aleada/Qwen3.6-27B-W4A16
Image-Text-to-Text • 28B • Updated • 76Note W4A16 for Ampere Marlin kernels, with the MTP head grafted back and exempted from the quantization config. Ships a measured draft-acceptance rate (86.5%, 1.73x decode) rather than an assumption - other 4-bit builds of this model retain the head but none publish what it actually accepts.