ConvNeXt (fastforks)

ConvNeXt (Liu et al. 2022): a pure convolutional network of 7x7 depthwise conv + LayerNorm + MLP blocks, Meta's ImageNet-1k and ImageNet-22k checkpoints.

This repo is the model card only: it holds no weights. fastforks downloads the authors' weights from the links below and checks them against the SHA-256 in the model card. Code: https://github.com/kleinicke/fastforks. Kernels for the kernels backend: https://huggingface.co/fastforks/kernels. Original code: https://github.com/huggingface/pytorch-image-models at 0df212b369a5.

Author weight repos: https://huggingface.co/timm/convnext_base.fb_in1k, https://huggingface.co/timm/convnext_base.fb_in22k_ft_in1k, https://huggingface.co/timm/convnext_base.fb_in22k_ft_in1k_384, https://huggingface.co/timm/convnext_large.fb_in1k, https://huggingface.co/timm/convnext_large.fb_in22k_ft_in1k, https://huggingface.co/timm/convnext_large.fb_in22k_ft_in1k_384, https://huggingface.co/timm/convnext_small.fb_in1k, https://huggingface.co/timm/convnext_small.fb_in22k_ft_in1k, https://huggingface.co/timm/convnext_small.fb_in22k_ft_in1k_384, https://huggingface.co/timm/convnext_tiny.fb_in1k, https://huggingface.co/timm/convnext_tiny.fb_in22k_ft_in1k, https://huggingface.co/timm/convnext_tiny.fb_in22k_ft_in1k_384.

Task: classification. Tier: stable. Outputs: logits. Backends: portable, kernels.

Input: one RGB image. Output: ImageNet-1k class logits.

Variants

Weights are the authors' files (not hosted here). SHA-256 shortened to 12 hex digits; the full hashes are in card.toml.

variant weights (author link) SHA-256 weights licence commercial use
convnext_base.fb_in1k@1 convnext_base.fb_in1k.safetensors 0d397521d594 Apache-2.0 yes
convnext_base.fb_in22k_ft_in1k@1 convnext_base.fb_in22k_ft_in1k.safetensors 4853efc8d5dd Apache-2.0 yes
convnext_base.fb_in22k_ft_in1k_384@1 convnext_base.fb_in22k_ft_in1k_384.safetensors bc0cfafc7d77 Apache-2.0 yes
convnext_large.fb_in1k@1 convnext_large.fb_in1k.safetensors 657d0be988c2 Apache-2.0 yes
convnext_large.fb_in22k_ft_in1k@1 convnext_large.fb_in22k_ft_in1k.safetensors e70c6c68fbd2 Apache-2.0 yes
convnext_large.fb_in22k_ft_in1k_384@1 convnext_large.fb_in22k_ft_in1k_384.safetensors d2d9196af67d Apache-2.0 yes
convnext_small.fb_in1k@1 convnext_small.fb_in1k.safetensors 5ec5ed9e97fb Apache-2.0 yes
convnext_small.fb_in22k_ft_in1k@1 convnext_small.fb_in22k_ft_in1k.safetensors 2457488bec2f Apache-2.0 yes
convnext_small.fb_in22k_ft_in1k_384@1 convnext_small.fb_in22k_ft_in1k_384.safetensors 17e4aa3c6ba4 Apache-2.0 yes
convnext_tiny.fb_in1k@1 convnext_tiny.fb_in1k.safetensors 08b9dc9c3a3a Apache-2.0 yes
convnext_tiny.fb_in22k_ft_in1k@1 convnext_tiny.fb_in22k_ft_in1k.safetensors 77b619e187de Apache-2.0 yes
convnext_tiny.fb_in22k_ft_in1k_384@1 convnext_tiny.fb_in22k_ft_in1k_384.safetensors a5cbe10cfa2e Apache-2.0 yes

Licence

Original-code strings, recorded as they are (research-use package; check them before any other use). Code: Apache-2.0.

Training data: ImageNet-21k / ImageNet-22k (full ImageNet, fall 2011 release), ImageNet-1k (ILSVRC 2012 classification).

Licence evidence: https://huggingface.co/timm/convnext_base.fb_in22k_ft_in1k/blob/67758fcbcf9f007ec7630bf8636b74a8f24e52a0/README.md

Input size

  • convnext_base.fb_in1k, convnext_base.fb_in22k_ft_in1k, convnext_large.fb_in1k, convnext_large.fb_in22k_ft_in1k, convnext_small.fb_in1k, convnext_small.fb_in22k_ft_in1k, convnext_tiny.fb_in1k, convnext_tiny.fb_in22k_ft_in1k: fixed 224x224: the input is resized (short side 256) and centre-cropped to it
  • convnext_base.fb_in22k_ft_in1k_384, convnext_large.fb_in22k_ft_in1k_384, convnext_small.fb_in22k_ft_in1k_384, convnext_tiny.fb_in22k_ft_in1k_384: fixed 384x384: the input is resized (short side 384) and centre-cropped to it Tested up to (portable, fp16, RTX 5060 Ti 16 GB): convnext_base.fb_in1k 224x224 (0.3 GiB peak allocated, 0.4 reserved); convnext_base.fb_in22k_ft_in1k 224x224 (0.3 GiB peak allocated, 0.4 reserved); convnext_base.fb_in22k_ft_in1k_384 384x384 (0.3 GiB peak allocated, 0.4 reserved); convnext_large.fb_in1k 224x224 (0.6 GiB peak allocated, 0.7 reserved); convnext_large.fb_in22k_ft_in1k 224x224 (0.6 GiB peak allocated, 0.7 reserved); convnext_large.fb_in22k_ft_in1k_384 384x384 (0.6 GiB peak allocated, 0.7 reserved); convnext_small.fb_in1k 224x224 (0.2 GiB peak allocated, 0.2 reserved); convnext_small.fb_in22k_ft_in1k 224x224 (0.2 GiB peak allocated, 0.2 reserved); convnext_small.fb_in22k_ft_in1k_384 384x384 (0.2 GiB peak allocated, 0.2 reserved); convnext_tiny.fb_in1k 224x224 (0.1 GiB peak allocated, 0.2 reserved); convnext_tiny.fb_in22k_ft_in1k 224x224 (0.1 GiB peak allocated, 0.2 reserved); convnext_tiny.fb_in22k_ft_in1k_384 384x384 (0.1 GiB peak allocated, 0.2 reserved)

Speed and memory

Classification, synthetic

model input time (ms): original fp32 original portable kernels kernels speed-up: vs original vs portable peak memory (MiB, allocated / reserved): original portable kernels
convnext_base.fb_in1k 224x224 7.77 11.2 5.57 2.01 3.9x 2.8x 541 / 566 291 / 296 208 / 236
convnext_base.fb_in22k_ft_in1k 224x224 7.78 11.3 5.66 2.00 3.9x 2.8x 541 / 566 291 / 296 208 / 236
convnext_base.fb_in22k_ft_in1k_384 224x224 11.6 11.2 5.48 3.00 3.7x 1.8x 549 / 562 309 / 334 213 / 232
convnext_large.fb_in1k 224x224 10.7 11.5 5.59 3.29 3.2x 1.7x 1166 / 1180 603 / 632 419 / 434
convnext_large.fb_in22k_ft_in1k 224x224 10.7 11.5 5.57 3.28 3.2x 1.7x 1166 / 1180 603 / 632 419 / 434
convnext_large.fb_in22k_ft_in1k_384 224x224 22.9 11.5 5.99 5.20 2.2x 1.2x 1177 / 1260 629 / 654 426 / 454
convnext_small.fb_in1k 224x224 7.44 11.3 5.69 1.84 4.0x 3.1x 321 / 338 181 / 190 134 / 152
convnext_small.fb_in22k_ft_in1k 224x224 7.51 11.3 5.73 1.84 4.1x 3.1x 321 / 338 181 / 190 134 / 152
convnext_small.fb_in22k_ft_in1k_384 224x224 8.04 11.2 5.51 2.52 3.2x 2.2x 328 / 338 194 / 212 138 / 152
convnext_tiny.fb_in1k 224x224 4.05 6.14 3.20 1.15 3.5x 2.8x 198 / 216 119 / 130 92 / 110
convnext_tiny.fb_in22k_ft_in1k 224x224 4.11 6.13 3.21 1.15 3.6x 2.8x 198 / 216 119 / 130 92 / 110
convnext_tiny.fb_in22k_ft_in1k_384 224x224 4.59 6.05 3.15 1.56 2.9x 2.0x 204 / 216 133 / 152 96 / 110

Startup and first call

model load (s): original portable kernels first call (s): original portable kernels
convnext_base.fb_in1k 0.95 1.91 1.99 0.01 0.25 0.27
convnext_base.fb_in22k_ft_in1k 0.89 2.07 1.97 0.01 0.25 0.26
convnext_base.fb_in22k_ft_in1k_384 0.87 2.54 1.91 0.01 0.25 0.26
convnext_large.fb_in1k 1.89 2.79 2.58 0.02 0.25 0.24
convnext_large.fb_in22k_ft_in1k 1.91 2.55 2.62 0.01 0.25 0.28
convnext_large.fb_in22k_ft_in1k_384 1.75 2.58 2.51 0.02 0.25 0.25
convnext_small.fb_in1k 0.65 1.81 1.59 0.01 0.22 0.19
convnext_small.fb_in22k_ft_in1k 0.52 1.67 1.93 0.01 0.22 0.21
convnext_small.fb_in22k_ft_in1k_384 0.52 1.70 1.81 0.01 0.22 0.36
convnext_tiny.fb_in1k 0.41 1.48 1.50 0.49 0.24 0.24
convnext_tiny.fb_in22k_ft_in1k 0.35 1.49 1.45 0.01 0.24 0.24
convnext_tiny.fb_in22k_ft_in1k_384 0.35 1.57 1.48 0.01 0.24 0.24

torch.compile and TensorRT

model input time: portable compile TensorRT fp16 kernels preparation: compile (cold first call) ONNX + engine build peak MiB (alloc / reserved): portable compile kernels TensorRT (engine + torch)
convnext_base.fb_in1k 224x224 5.57 ms 3.37 ms 1.45 ms 2.01 ms 10.9 s 32.5 s 291 / 296 323 / 334 208 / 236 482 / 494
convnext_base.fb_in22k_ft_in1k 224x224 5.66 ms 3.37 ms 1.45 ms 2.00 ms 11.4 s 31.3 s 291 / 296 323 / 334 208 / 236 482 / 494
convnext_base.fb_in22k_ft_in1k_384 224x224 5.48 ms 3.41 ms 2.31 ms 3.00 ms 10.8 s 31.2 s 309 / 334 336 / 354 213 / 232 493 / 522
convnext_large.fb_in1k 224x224 5.59 ms 3.45 ms 2.50 ms 3.29 ms 10.8 s 33.9 s 603 / 632 642 / 652 419 / 434 986 / 1006
convnext_large.fb_in22k_ft_in1k 224x224 5.57 ms 3.44 ms 2.50 ms 3.28 ms 10.8 s 35.0 s 603 / 632 642 / 652 419 / 434 986 / 1006
convnext_large.fb_in22k_ft_in1k_384 224x224 5.99 ms 5.80 ms 4.23 ms 5.20 ms 10.7 s 33.4 s 629 / 654 654 / 672 426 / 454 1001 / 1040
convnext_small.fb_in1k 224x224 5.69 ms 3.47 ms 1.16 ms 1.84 ms 10.6 s 30.5 s 181 / 190 211 / 230 134 / 152 278 / 294
convnext_small.fb_in22k_ft_in1k 224x224 5.73 ms 3.50 ms 1.17 ms 1.84 ms 10.5 s 30.9 s 181 / 190 211 / 230 134 / 152 278 / 294
convnext_small.fb_in22k_ft_in1k_384 224x224 5.51 ms 3.36 ms 1.68 ms 2.52 ms 10.7 s 31.8 s 194 / 212 223 / 232 138 / 152 317 / 350
convnext_tiny.fb_in1k 224x224 3.20 ms 2.09 ms 0.682 ms 1.15 ms 8.22 s 27.4 s 119 / 130 150 / 170 92 / 110 184 / 200
convnext_tiny.fb_in22k_ft_in1k 224x224 3.21 ms 2.08 ms 0.687 ms 1.15 ms 8.09 s 28.3 s 119 / 130 150 / 170 92 / 110 184 / 200
convnext_tiny.fb_in22k_ft_in1k_384 224x224 3.15 ms 2.05 ms 0.970 ms 1.56 ms 7.90 s 28.2 s 133 / 152 162 / 170 96 / 110 192 / 226

Fixed-size cells: time (ms) / peak allocated / peak reserved memory (MiB).

Hardware, protocol and caveats: benchmarks (RTX 5060 Ti 16 GB, sm_120). Original = the original code named above at the precision of portable (fp16, or (fp32) where fp16 is unsupported or not finite); original fp32 = its default fp32 run. Peak memory is the torch allocator's.

Accuracy

On public datasets

fastforks: the model's own resize / crop on the GPU; original: the library's pretrained loader and its PIL eval transform. Original = the original code named under Details, same checkpoint and protocol.

variant dataset original fp32 original fp16 portable kernels note
convnext_tiny.fb_in1k ImageNet-1k val, 5000 random images, top-1 / top-5 % 82.32 / 96.40 82.28 / 96.40 82.34 / 96.28 82.28 / 96.32
convnext_small.fb_in1k ImageNet-1k val, 5000 random images, top-1 / top-5 % 83.66 / 96.64 83.68 / 96.64 83.64 / 96.56 83.66 / 96.56
convnext_base.fb_in1k ImageNet-1k val, 5000 random images, top-1 / top-5 % 84.66 / 96.88 84.64 / 96.86 84.58 / 96.82 84.54 / 96.82
convnext_large.fb_in1k ImageNet-1k val, 5000 random images, top-1 / top-5 % 85.44 / 96.78 85.40 / 96.78 85.54 / 96.78 85.56 / 96.84
convnext_tiny.fb_in22k_ft_in1k ImageNet-1k val, 5000 random images, top-1 / top-5 % 83.64 / 96.82 83.66 / 96.82 83.46 / 96.78 83.44 / 96.80
convnext_small.fb_in22k_ft_in1k ImageNet-1k val, 5000 random images, top-1 / top-5 % 84.78 / 97.66 84.76 / 97.66 84.96 / 97.66 84.96 / 97.66
convnext_base.fb_in22k_ft_in1k ImageNet-1k val, 5000 random images, top-1 / top-5 % 86.42 / 97.96 86.48 / 97.96 86.26 / 97.94 86.30 / 97.94
convnext_large.fb_in22k_ft_in1k ImageNet-1k val, 5000 random images, top-1 / top-5 % 87.54 / 98.26 87.54 / 98.26 87.38 / 98.26 87.38 / 98.26
convnext_tiny.fb_in22k_ft_in1k_384 ImageNet-1k val, 5000 random images, top-1 / top-5 % 84.98 / 97.40 84.96 / 97.40 85.00 / 97.22 84.94 / 97.22
convnext_small.fb_in22k_ft_in1k_384 ImageNet-1k val, 5000 random images, top-1 / top-5 % 86.36 / 98.04 86.36 / 98.04 86.16 / 98.18 86.16 / 98.18
convnext_base.fb_in22k_ft_in1k_384 ImageNet-1k val, 5000 random images, top-1 / top-5 % 87.48 / 98.50 87.52 / 98.50 87.34 / 98.40 87.32 / 98.40
convnext_large.fb_in22k_ft_in1k_384 ImageNet-1k val, 5000 random images, top-1 / top-5 % 88.36 / 98.50 88.34 / 98.50 88.08 / 98.50 88.08 / 98.50

Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-10.

Parity with the original code

Every variant is checked against reference outputs of the original code at its pinned commit. CPU, float32: max |diff| <= 1.7e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.0033 and p99 <= 0.012 (absolute error of the output logits).

Usage

from fastforks import classification

model = classification.load("convnext_base.fb_in1k")  # mode="auto": the preferred backend from recorded measurements
result = model("photo.jpg")
result.top(5)  # top-5 classes
result.backend  # what ran, e.g. Backend(mode='portable', device='cuda:0', dtype='float16')

Install with pip install fastforks (see https://github.com/kleinicke/fastforks). Other variants: convnext_base.fb_in1k, convnext_base.fb_in22k_ft_in1k, convnext_base.fb_in22k_ft_in1k_384, convnext_large.fb_in1k, convnext_large.fb_in22k_ft_in1k, convnext_large.fb_in22k_ft_in1k_384, convnext_small.fb_in1k, convnext_small.fb_in22k_ft_in1k, convnext_small.fb_in22k_ft_in1k_384, convnext_tiny.fb_in1k, convnext_tiny.fb_in22k_ft_in1k, convnext_tiny.fb_in22k_ft_in1k_384.

Kernels backend

classification.load("convnext_base.fb_in1k", mode="kernels") runs our CUDA kernels (fvm_conv, fvm_seg). It needs a prebuilt kernel for your GPU architecture (downloaded from https://huggingface.co/fastforks/kernels on first use) and a recent NVIDIA driver; if there is none it raises BackendUnavailableError with the reason and never falls back silently. mode="auto" picks kernels when it can run and logs the choice.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Collection including fastforks/convnext