Models · Classification

ConvNeXt

Meta's ConvNeXt (Liu et al. 2022; tiny, small, base, large; ImageNet-1k, and ImageNet-22k fine-tuned at 224 and 384): a convolutional classifier of 7x7 depthwise convolutions, LayerNorm and MLP blocks. Apache-2.0.

stable Fork of huggingface/pytorch-image-models@0df212b · backends: portable, kernels

  • Paper A ConvNet for the 2020s CVPR 2022 · first released 10 Jan 2022
  • Authors’ code facebookresearch/ConvNeXt (PyTorch), not the code forked here: timm ships a maintained re-implementation with the same Meta ImageNet-1k/22k weights and a uniform API.
Checkpoint

Paper 2022 Added in 0.2.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 2.01 ms 224x224, fp16
vs the faster original 3.9× original fp32: 7.78 ms
Peak memory 208 MiB original: 390 MiB
Accuracy vs original fp32 −0.12 kernels top-1 %, ImageNet-1k val (86.42 original)

Use it

pip install fastforks

Lean install with only what this model needs: pip install "fastforks-core[timm]"

from fastforks import classification

model = classification.load("convnext_base.fb_in22k_ft_in1k")  # kernels on a supported GPU, else portable PyTorch
result = model(image)
result.logits
result.backend  # what ran: mode, device, dtype, kernels
fastforks bench convnext_base.fb_in22k_ft_in1k

Times the model on this machine. load(…, mode="portable") or mode="kernels" picks a backend; usage guide.

Speed and memory

Every backend on one GPU (RTX 5060 Ti 16 GB) and the same input, the authors’ code included. Kernels runs fp16. Portable runs fp16 on CUDA.

Input: synthetic, 224x224 (the task’s standard size)

BackenddtypeTime per call (ms)vs original fp32Peak memory
allocated / reserved
Load (s)First call (s)
Original fp32timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed callfp32 7.78–390 MiB/398 MiB1.00.01
Original fp16torch.autocast(fp16) around the call; timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed callfp16 11.30.69×541 MiB/566 MiB0.90.01
Portable fp32fp32not measured ––––
Portablefp16 5.841.3×211 MiB/234 MiB2.10.27
Kernelsfp16 2.013.9×208 MiB/236 MiB2.40.26
Portable + torch.compile 3.372.3×323 MiB/334 MiB11 (build)–
TensorRT fp16fp16 1.455.4×200 MiBenginenot comparable31 (build)–

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-10 to 2026-11-16. First call, 3 warm-up, peak reset, median of 3 rounds x median of 5. Ratios divide two times of this table (same GPU, same input). Original fp16 is the authors’ own precision option, else torch.autocast. torch.compile and TensorRT: build time instead of load time. TensorRT memory (engine, not comparable) is the GPU memory the engine took, read from the driver because TensorRT allocates outside PyTorch, so it can't be compared with the PyTorch peaks of the other rows.

Other devices

DeviceBackenddtypems / callPeak memoryOutput vs referenceInput
Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-14Originalfp3225.7 2,092 MiB process–224x224
Portablefp1624.9 2,308 MiB processmatches224x224
Portablefp3227.8 2,324 MiB processmatches224x224
Kernelsfp16not implemented–224x224

CUDA memory is torch’s peak allocated / reserved; Apple has no such counter, so the Mac rows show the process footprint (weights, Python and torch included). “Output vs reference” compares the device’s output with the stored golden output under the card’s tolerance; – = not checked on that device. Kernels “not implemented” = no kernels for this model on that backend yet: it runs portable there.

Licence and weights

Commercial use allowed

Code
Apache-2.0
Weights
Apache-2.0 source
Training data

Our reading of the licences, not legal advice.

Weights

convnext_base.fb_in22k_ft_in1k.safetensors 354 MB· hosted on Hugging Face

fastforks downloads and checks these files on first use (fastforks download convnext_base.fb_in22k_ft_in1k).

Accuracy

The test set through the authors’ code and through ours: same checkpoint, same protocol. The distance between the original’s fp32 and fp16 runs is what half precision alone costs; our runs should stay within it.

ImageNet-1k val

5000 random images, top-1 / top-5 %

Rundtypetop-1 % ↑top-5 % ↑Δ top-1 % vs fp32 originalimagesms / image
Original fp32fp3286.4297.96 reference 5000 3.87
Original fp16fp1686.4897.96 +0.06 5000 1.73
Portable fp32fp3286.3497.94 −0.08 5000 4.22
Portablefp1686.2697.94 −0.16 5000 1.18
Kernelsfp1686.3097.94 −0.12 5000 0.943

fastforks: the model's own resize / crop on the GPU; original: the library's pretrained loader and its PIL eval transform. Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16, portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-10; portable fp32: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 6f0a70e1, 2026-11-17.

Closeness to the original

Beyond the test-set scores: the raw output of ours against the original’s fp32 output on the same input, pixel by pixel. 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).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.000017 not recorded
GPU default dtypes, mean |diff|≤ 0.0033 0.00099
GPU default dtypes, p99 |diff|≤ 0.012 0.00361

Recorded for convnext_base.fb_in22k_ft_in1k on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-10, with these kernels: fvm_convfvm_seg

Details

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

Input / output
Input: one RGB image. Output: ImageNet-1k class logits.
Input size
fixed 224x224: the input is resized (short side 256) and centre-cropped to it; tested up to 224x224 (291 MiB allocated)
Outputs
logits
Checkpoint
arch="convnext_base.fb_in22k_ft_in1k"
CUDA kernels
  • fvm_convfused convolutions for image classifiers
  • fvm_segfused kernels for Mask2Former and OneFormer
Apple (MPS)
fp32, fp16
Full reference, generated from the repository

Meta's ConvNeXt (Liu et al. 2022; tiny, small, base, large; ImageNet-1k, and ImageNet-22k fine-tuned at 224 and 384): a convolutional classifier of 7x7 depthwise convolutions, LayerNorm and MLP blocks. Apache-2.0.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
convnext_tiny.fb_in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
4.05
6.14
3.31
1.15
2.09
0.682
155 / 170
198 / 216
94 / 110
92 / 110
150 / 170
184 / 200
convnext_small.fb_in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
7.44
11.3
5.97
1.84
3.47
1.16
238 / 252
321 / 338
136 / 150
134 / 152
211 / 230
278 / 294
convnext_base.fb_in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
7.77
11.2
5.84
2.00
3.37
1.45
390 / 398
541 / 566
211 / 234
208 / 236
323 / 334
482 / 494
convnext_large.fb_in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
10.7
11.5
5.81
3.28
3.45
2.50
828 / 854
1166 / 1180
423 / 432
419 / 434
642 / 652
986 / 1006
convnext_tiny.fb_in22k_ft_in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
4.11
6.13
3.34
1.15
2.08
0.687
155 / 172
198 / 216
94 / 110
92 / 110
150 / 170
184 / 200
convnext_small.fb_in22k_ft_in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
7.51
11.3
5.84
1.84
3.50
1.17
238 / 252
321 / 338
136 / 150
134 / 152
211 / 230
278 / 294
convnext_base.fb_in22k_ft_in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
7.78
11.3
5.84
2.01
3.37
1.45
390 / 398
541 / 566
211 / 234
208 / 236
323 / 334
482 / 494
convnext_large.fb_in22k_ft_in1ksynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
10.7
11.5
5.88
3.28
3.44
2.50
828 / 854
1166 / 1180
423 / 432
419 / 434
642 / 652
986 / 1006
convnext_tiny.fb_in22k_ft_in1k_384synthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
4.59
6.05
3.30
1.56
2.05
0.970
182 / 210
204 / 216
108 / 132
96 / 110
162 / 170
192 / 226
convnext_small.fb_in22k_ft_in1k_384synthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
8.04
11.2
5.82
2.53
3.36
1.68
264 / 292
328 / 338
150 / 172
138 / 152
223 / 232
317 / 350
convnext_base.fb_in22k_ft_in1k_384synthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
11.6
11.2
5.79
3.05
3.41
2.31
423 / 436
549 / 562
229 / 252
213 / 232
336 / 354
493 / 522
convnext_large.fb_in22k_ft_in1k_384synthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
22.9
11.5
6.11
5.21
5.80
4.23
864 / 894
1177 / 1260
449 / 492
426 / 454
654 / 672
1001 / 1040
Details
  • Original code: https://github.com/huggingface/pytorch-image-models at 0df212b369a5
  • Code licence: Apache-2.0 (permissive); files:
  • Weights licence: Apache-2.0
  • Licence evidence: https://huggingface.co/timm/convnext_base.fb_in22k_ft_in1k/blob/67758fcbcf9f007ec7630bf8636b74a8f24e52a0/README.md
  • Training data: ImageNet-21k / ImageNet-22k (full ImageNet, fall 2011 release), ImageNet-1k (ILSVRC 2012 classification)
    • convnext_tiny.fb_in1k: ImageNet-1k (ILSVRC 2012 classification)
    • convnext_small.fb_in1k: ImageNet-1k (ILSVRC 2012 classification)
    • convnext_base.fb_in1k: ImageNet-1k (ILSVRC 2012 classification)
    • convnext_large.fb_in1k: ImageNet-1k (ILSVRC 2012 classification)
  • Kernel libraries: fvm_conv, fvm_seg
  • Variants: convnext_tiny.fb_in1k, convnext_small.fb_in1k, convnext_base.fb_in1k, convnext_large.fb_in1k, convnext_tiny.fb_in22k_ft_in1k, convnext_small.fb_in22k_ft_in1k, convnext_base.fb_in22k_ft_in1k (default of convnext_base), convnext_large.fb_in22k_ft_in1k (default of convnext_large), convnext_tiny.fb_in22k_ft_in1k_384, convnext_small.fb_in22k_ft_in1k_384, convnext_base.fb_in22k_ft_in1k_384, convnext_large.fb_in22k_ft_in1k_384
  • Outputs: logits; label_set: imagenet-1k
  • Input size:
    • convnext_tiny.fb_in1k, convnext_small.fb_in1k, convnext_base.fb_in1k, convnext_large.fb_in1k, convnext_tiny.fb_in22k_ft_in1k, convnext_small.fb_in22k_ft_in1k, convnext_base.fb_in22k_ft_in1k, convnext_large.fb_in22k_ft_in1k: fixed 224x224: the input is resized (short side 256) and centre-cropped to it
    • convnext_tiny.fb_in22k_ft_in1k_384, convnext_small.fb_in22k_ft_in1k_384, convnext_base.fb_in22k_ft_in1k_384, convnext_large.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_tiny.fb_in1k 224x224 (0.1 GiB peak allocated, 0.2 reserved); convnext_small.fb_in1k 224x224 (0.2 GiB peak allocated, 0.2 reserved); convnext_base.fb_in1k 224x224 (0.3 GiB peak allocated, 0.4 reserved); convnext_large.fb_in1k 224x224 (0.6 GiB peak allocated, 0.7 reserved); convnext_tiny.fb_in22k_ft_in1k 224x224 (0.1 GiB peak allocated, 0.2 reserved); convnext_small.fb_in22k_ft_in1k 224x224 (0.2 GiB peak allocated, 0.2 reserved); convnext_base.fb_in22k_ft_in1k 224x224 (0.3 GiB peak allocated, 0.4 reserved); convnext_large.fb_in22k_ft_in1k 224x224 (0.6 GiB peak allocated, 0.7 reserved); convnext_tiny.fb_in22k_ft_in1k_384 384x384 (0.1 GiB peak allocated, 0.2 reserved); convnext_small.fb_in22k_ft_in1k_384 384x384 (0.2 GiB peak allocated, 0.2 reserved); convnext_base.fb_in22k_ft_in1k_384 384x384 (0.3 GiB peak allocated, 0.4 reserved); convnext_large.fb_in22k_ft_in1k_384 384x384 (0.6 GiB peak allocated, 0.7 reserved)
  • convnext_tiny.fb_in1k 224x224, first call: portable + torch.compile 8 s (cold Inductor cache), TensorRT fp16 27 s (ONNX + engine).
  • convnext_small.fb_in1k 224x224, first call: portable + torch.compile 11 s (cold Inductor cache), TensorRT fp16 30 s (ONNX + engine).
  • convnext_base.fb_in1k 224x224, first call: portable + torch.compile 11 s (cold Inductor cache), TensorRT fp16 32 s (ONNX + engine).
  • convnext_large.fb_in1k 224x224, first call: portable + torch.compile 11 s (cold Inductor cache), TensorRT fp16 34 s (ONNX + engine).
  • convnext_tiny.fb_in22k_ft_in1k 224x224, first call: portable + torch.compile 8 s (cold Inductor cache), TensorRT fp16 28 s (ONNX + engine).
  • convnext_small.fb_in22k_ft_in1k 224x224, first call: portable + torch.compile 10 s (cold Inductor cache), TensorRT fp16 31 s (ONNX + engine).
  • convnext_base.fb_in22k_ft_in1k 224x224, first call: portable + torch.compile 11 s (cold Inductor cache), TensorRT fp16 31 s (ONNX + engine).
  • convnext_large.fb_in22k_ft_in1k 224x224, first call: portable + torch.compile 11 s (cold Inductor cache), TensorRT fp16 35 s (ONNX + engine).
  • convnext_tiny.fb_in22k_ft_in1k_384 224x224, first call: portable + torch.compile 8 s (cold Inductor cache), TensorRT fp16 28 s (ONNX + engine).
  • convnext_small.fb_in22k_ft_in1k_384 224x224, first call: portable + torch.compile 11 s (cold Inductor cache), TensorRT fp16 32 s (ONNX + engine).
  • convnext_base.fb_in22k_ft_in1k_384 224x224, first call: portable + torch.compile 11 s (cold Inductor cache), TensorRT fp16 31 s (ONNX + engine).
  • convnext_large.fb_in22k_ft_in1k_384 224x224, first call: portable + torch.compile 11 s (cold Inductor cache), TensorRT fp16 33 s (ONNX + engine).
  • convnext_tiny.fb_in1k original (fp32) 224x224: timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_tiny.fb_in1k original 224x224: torch.autocast(fp16) around the call; timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_small.fb_in1k original (fp32) 224x224: timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_small.fb_in1k original 224x224: torch.autocast(fp16) around the call; timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_base.fb_in1k original (fp32) 224x224: timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_base.fb_in1k original 224x224: torch.autocast(fp16) around the call; timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_large.fb_in1k original (fp32) 224x224: timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_large.fb_in1k original 224x224: torch.autocast(fp16) around the call; timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_tiny.fb_in22k_ft_in1k original (fp32) 224x224: timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_tiny.fb_in22k_ft_in1k original 224x224: torch.autocast(fp16) around the call; timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_small.fb_in22k_ft_in1k original (fp32) 224x224: timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_small.fb_in22k_ft_in1k original 224x224: torch.autocast(fp16) around the call; timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_base.fb_in22k_ft_in1k original (fp32) 224x224: timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_base.fb_in22k_ft_in1k original 224x224: torch.autocast(fp16) around the call; timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_large.fb_in22k_ft_in1k original (fp32) 224x224: timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_large.fb_in22k_ft_in1k original 224x224: torch.autocast(fp16) around the call; timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_tiny.fb_in22k_ft_in1k_384 original (fp32) 224x224: timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_tiny.fb_in22k_ft_in1k_384 original 224x224: torch.autocast(fp16) around the call; timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_small.fb_in22k_ft_in1k_384 original (fp32) 224x224: timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_small.fb_in22k_ft_in1k_384 original 224x224: torch.autocast(fp16) around the call; timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_base.fb_in22k_ft_in1k_384 original (fp32) 224x224: timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_base.fb_in22k_ft_in1k_384 original 224x224: torch.autocast(fp16) around the call; timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_large.fb_in22k_ft_in1k_384 original (fp32) 224x224: timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_large.fb_in22k_ft_in1k_384 original 224x224: torch.autocast(fp16) around the call; timm 1.0.30 network, bicubic resize + centre crop + normalisation on the GPU inside the timed call
  • convnext_tiny.fb_in1k load / first call: original 0.41 s / 0.491 s, portable 1.48 s / 0.242 s, kernels 1.5 s / 0.245 s
  • convnext_small.fb_in1k load / first call: original 0.65 s / 0.014 s, portable 1.81 s / 0.219 s, kernels 1.59 s / 0.186 s
  • convnext_base.fb_in1k load / first call: original 0.95 s / 0.015 s, portable 1.91 s / 0.25 s, kernels 1.99 s / 0.27 s
  • convnext_large.fb_in1k load / first call: original 1.89 s / 0.016 s, portable 2.79 s / 0.251 s, kernels 2.58 s / 0.244 s
  • convnext_tiny.fb_in22k_ft_in1k load / first call: original 0.35 s / 0.008 s, portable 1.49 s / 0.241 s, kernels 1.45 s / 0.237 s
  • convnext_small.fb_in22k_ft_in1k load / first call: original 0.52 s / 0.013 s, portable 1.67 s / 0.222 s, kernels 1.93 s / 0.206 s
  • convnext_base.fb_in22k_ft_in1k load / first call: original 0.89 s / 0.014 s, portable 2.07 s / 0.251 s, kernels 1.97 s / 0.261 s
  • convnext_large.fb_in22k_ft_in1k load / first call: original 1.91 s / 0.015 s, portable 2.55 s / 0.25 s, kernels 2.62 s / 0.278 s
  • convnext_tiny.fb_in22k_ft_in1k_384 load / first call: original 0.35 s / 0.01 s, portable 1.57 s / 0.243 s, kernels 1.48 s / 0.242 s
  • convnext_small.fb_in22k_ft_in1k_384 load / first call: original 0.52 s / 0.013 s, portable 1.7 s / 0.217 s, kernels 1.81 s / 0.364 s
  • convnext_base.fb_in22k_ft_in1k_384 load / first call: original 0.87 s / 0.014 s, portable 2.54 s / 0.248 s, kernels 1.91 s / 0.257 s
  • convnext_large.fb_in22k_ft_in1k_384 load / first call: original 1.75 s / 0.016 s, portable 2.58 s / 0.253 s, kernels 2.51 s / 0.251 s
  • Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5

Measured

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.

variantdatasetoriginal fp32original fp16portableportable fp32kernelsnote
convnext_tiny.fb_in1kImageNet-1k val, 5000 random images, top-1 / top-5 %82.32 / 96.4082.28 / 96.4082.34 / 96.2882.30 / 96.2882.28 / 96.32
convnext_small.fb_in1kImageNet-1k val, 5000 random images, top-1 / top-5 %83.66 / 96.6483.68 / 96.6483.64 / 96.5683.62 / 96.5683.66 / 96.56
convnext_base.fb_in1kImageNet-1k val, 5000 random images, top-1 / top-5 %84.66 / 96.8884.64 / 96.8684.58 / 96.8284.56 / 96.8084.54 / 96.82
convnext_large.fb_in1kImageNet-1k val, 5000 random images, top-1 / top-5 %85.44 / 96.7885.40 / 96.7885.54 / 96.7885.56 / 96.8085.56 / 96.84
convnext_tiny.fb_in22k_ft_in1kImageNet-1k val, 5000 random images, top-1 / top-5 %83.64 / 96.8283.66 / 96.8283.46 / 96.7883.46 / 96.8083.44 / 96.80
convnext_small.fb_in22k_ft_in1kImageNet-1k val, 5000 random images, top-1 / top-5 %84.78 / 97.6684.76 / 97.6684.96 / 97.6684.98 / 97.6684.96 / 97.66
convnext_base.fb_in22k_ft_in1kImageNet-1k val, 5000 random images, top-1 / top-5 %86.42 / 97.9686.48 / 97.9686.26 / 97.9486.34 / 97.9486.30 / 97.94
convnext_large.fb_in22k_ft_in1kImageNet-1k val, 5000 random images, top-1 / top-5 %87.54 / 98.2687.54 / 98.2687.38 / 98.2687.36 / 98.2687.38 / 98.26
convnext_tiny.fb_in22k_ft_in1k_384ImageNet-1k val, 5000 random images, top-1 / top-5 %84.98 / 97.4084.96 / 97.4085.00 / 97.2284.94 / 97.2284.94 / 97.22
convnext_small.fb_in22k_ft_in1k_384ImageNet-1k val, 5000 random images, top-1 / top-5 %86.36 / 98.0486.36 / 98.0486.16 / 98.1886.14 / 98.1886.16 / 98.18
convnext_base.fb_in22k_ft_in1k_384ImageNet-1k val, 5000 random images, top-1 / top-5 %87.48 / 98.5087.52 / 98.5087.34 / 98.4087.32 / 98.4087.32 / 98.40
convnext_large.fb_in22k_ft_in1k_384ImageNet-1k val, 5000 random images, top-1 / top-5 %88.36 / 98.5088.34 / 98.5088.08 / 98.5088.06 / 98.5088.08 / 98.50

Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-10; portable32: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 6f0a70e1, 2026-11-17; kernels: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-10.

Parity

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).

variantkernels error vs original, mean / p99GPU, torchkernels used
convnext_tiny.fb_in1k0.00089 / 0.00297RTX 5060 Ti, torch 2.14.1+cu130fvm_conv@f1d064e539e1:sm_120, fvm_seg@739693dff3b4:sm_120
convnext_small.fb_in1k0.00062 / 0.00245RTX 5060 Ti, torch 2.14.1+cu130fvm_conv@f1d064e539e1:sm_120, fvm_seg@739693dff3b4:sm_120
convnext_base.fb_in1k0.00079 / 0.00318RTX 5060 Ti, torch 2.14.1+cu130fvm_conv@f1d064e539e1:sm_120, fvm_seg@739693dff3b4:sm_120
convnext_large.fb_in1k0.00071 / 0.00254RTX 5060 Ti, torch 2.14.1+cu130fvm_conv@f1d064e539e1:sm_120, fvm_seg@739693dff3b4:sm_120
convnext_tiny.fb_in22k_ft_in1k0.00215 / 0.00696RTX 5060 Ti, torch 2.14.1+cu130fvm_conv@f1d064e539e1:sm_120, fvm_seg@739693dff3b4:sm_120
convnext_small.fb_in22k_ft_in1k0.00086 / 0.00284RTX 5060 Ti, torch 2.14.1+cu130fvm_conv@f1d064e539e1:sm_120, fvm_seg@739693dff3b4:sm_120
convnext_base.fb_in22k_ft_in1k0.00099 / 0.00361RTX 5060 Ti, torch 2.14.1+cu130fvm_conv@f1d064e539e1:sm_120, fvm_seg@739693dff3b4:sm_120
convnext_large.fb_in22k_ft_in1k0.00111 / 0.00377RTX 5060 Ti, torch 2.14.1+cu130fvm_conv@f1d064e539e1:sm_120, fvm_seg@739693dff3b4:sm_120
convnext_tiny.fb_in22k_ft_in1k_3840.00216 / 0.00769RTX 5060 Ti, torch 2.14.1+cu130fvm_conv@f1d064e539e1:sm_120, fvm_seg@739693dff3b4:sm_120
convnext_small.fb_in22k_ft_in1k_3840.00141 / 0.00525RTX 5060 Ti, torch 2.14.1+cu130fvm_conv@f1d064e539e1:sm_120, fvm_seg@739693dff3b4:sm_120
convnext_base.fb_in22k_ft_in1k_3840.00102 / 0.00417RTX 5060 Ti, torch 2.14.1+cu130fvm_conv@f1d064e539e1:sm_120, fvm_seg@739693dff3b4:sm_120
convnext_large.fb_in22k_ft_in1k_3840.00107 / 0.00383RTX 5060 Ti, torch 2.14.1+cu130fvm_conv@f1d064e539e1:sm_120, fvm_seg@739693dff3b4:sm_120

Example output

One fixed input through this checkpoint, the same for every model of the task (kernels · float16 · NVIDIA GeForce RTX 5060 Ti).

Input

Top 5

  1. Egyptian cat 76%
  2. tiger cat 12%
  3. tabby 6.6%
  4. lynx 0.4%
  5. Angora <0.1%
Chelsea the cat, input
Input · 100%

Input: Chelsea the cat, Stéfan van der Walt (CC0).