Make AI fit your ideas.

ResNet-50Already pruned Vision 25.6 M params 49 MB fp16 A classic pruning benchmark with 50 convolutional layers. ViT-Large/16Already pruned Vision 307 M params 586 MB fp16 A large vision transformer using 16×16 image patches. DeiT-BaseAlready pruned Vision 86 M params 165 MB fp16 A vision transformer trained with distillation. MLP-Mixer-B/16Already pruned Vision 59 M params 113 MB fp16 An all-MLP vision backbone with 16×16 patches; FP16 size is estimated. DINOv3 ViT-B/16 Vision 86 M params 165 MB fp16 A self-supervised vision backbone. DINOv3 ViT-S/16 Vision 21 M params 40 MB fp16 The smaller DINOv3 variant, suited to constrained hardware. ConvNeXt-Tiny Vision 28 M params 54 MB fp16 A convolutional model with transformer-inspired design choices. YOLO26n Vision 2.4 M params 4.7 MB fp16 The nano variant, designed for on-device detection. YOLO26s Vision 9.5 M params 19 MB fp16 The small variant, balancing size and edge performance. YOLO26m Vision 20.4 M params 41 MB fp16 The medium variant, with higher memory needs on edge devices. RF-DETR-Nano Vision 30.5 M params 61 MB fp16 A detection transformer without anchors or NMS. DepthAnything v2 (Small) Vision 24.8 M params 47 MB fp16 Monocular depth estimation with a lightweight ViT-S backbone. DepthAnything v2 (Large) Vision 335.3 M params 640 MB fp16 The larger ViT-L variant of the same depth model family. DepthAnything v3 (Small) Vision 80 M params 160 MB fp16 A single transformer for depth and pose estimation. DepthAnything v3 (Base) Vision — params — fp16 The base v3 variant; reliable parameter counts are not yet available. SAM 3 Vision 830 M params 1.6 GB fp16 Segments objects from a text prompt. SegFormer-B0 Vision 3.7 M params 7.1 MB fp16 Lightweight semantic segmentation with a hierarchical encoder. CLIP ViT-B/32 Vision 151 M params 288 MB fp16 Maps text and images into one space for search and zero-shot tasks. Gemma 3 4B Language model 4.3 B params 8.6 GB fp16 Multimodal, with a 128K context window and modest memory needs. Gemma 3 12B Language model 12.2 B params 24.4 GB fp16 A larger variant that benefits from quantization on a single GPU. Gemma 3 27B Language model 27.4 B params 54.8 GB fp16 A high-capacity variant with substantial memory requirements. Qwen3 4B Language model 4.0 B params 8.0 GB fp16 A dense multilingual model released under Apache 2.0. Qwen3 8B Language model 8.2 B params 16.4 GB fp16 A practical mid-sized model for local use. Qwen3 30B-A3B Language model 30.5 B params 61 GB fp16 A mixture-of-experts model with about 3B active parameters per token. gpt-oss-20b Language model 21 B params 42 GB fp16 An Apache 2.0 mixture-of-experts model; quantization lowers memory use. DeepSeek-V4-Flash Language model 284 B params 568 GB fp16 A large mixture-of-experts model with about 13B active parameters. Request model + Can't find yours? Request it here.