March 23, 2026

• Improve pickle checkpoint handling security. Default all loading to weights_only=True, add safe_global for ArgParse.
• Improve attention mask handling for core ViT/EVA models & layers. Resolve bool masks, pass is_causal through for SSL tasks.
• Fix class & register token uses with ViT and no pos embed enabled.
• Add Patch Representation Refinement (PRR) as a pooling option in ViT. Thanks Sina (https://github.com/sinahmr).
• Improve consistency of output projection / MLP dimensions for attention pooling layers.
• Hiera model F.SDPA optimization to allow Flash Attention kernel use.
• Caution added to SGDP optimizer.
• Release 1.0.26. First maintenance release since my departure from Hugging Face.

What's Changed

• fix: replace 5 bare except clauses with except Exception by @haosenwang1018 in #2672
• Add timmx model export tool to README by @Boulaouaney in #2673
• Enhance SGDP optimizer with caution parameter by @Yuan-Jinghui in #2675
• Fix CLS and Reg tokens usage when pos_embed is disabled by @sinahmr in #2676
• default weights_only=True for load fns by @rwightman in #2679
• Fix Hiera global attention to use 4D tensors for efficient SDPA dispatch by @Raiden129 in #2680
• Improve 2d and latent attention pool dimension handling. Fix #2682 by @rwightman in #2684
• Improve attention mask handling for vision_transformer and eva and related blocks by @rwightman in #2686
• Implement PRR as a pooling module. Alternative to #2678 by @rwightman in #2685

New Contributors

• @haosenwang1018 made their first contribution in #2672
• @Raiden129 made their first contribution in #2680

Full Changelog: v1.0.25...v1.0.26

Feb 23, 2026

• Add token distillation training support to distillation task wrappers
• Remove some torch.jit usage in prep for official deprecation
• Caution added to AdamP optimizer
• Call reset_parameters() even if meta-device init so that buffers get init w/ hacks like init_empty_weights
• Tweak Muon optimizer to work with DTensor/FSDP2 (clamp_ instead of clamp_min_, alternate NS branch for DTensor)
• Release 1.0.25

Jan 21, 2026

• Compat Break: Fix oversight w/ QKV vs MLP bias in ParallelScalingBlock (& DiffParallelScalingBlock)
  • Does not impact any trained timm models but could impact downstream use.

What's Changed

• Token distill task & distill task refactoring by @rwightman in #2647
• Fix distilled head dropout using wrong token in PiT forward_head by @hassonofer in #2649
• Fix #2653, no models with weights impacted so just a clean fix by @rwightman in #2654
• Add the cautious optimizer to AdamP. by @Yuan-Jinghui in #2657
• Enhance the numerical stability of the Cautious Optimizer by @Yuan-Jinghui in #2658
• Some misc fixes for torch.jit deprecation and meta device init by @rwightman in #2664
• fix(optim): replace bare except with Exception in Lion optimizer by @llukito in #2666
• Change clamp_min_ to clamp_(min=) as former doesn't work with DTensor / FSDP2 by @rwightman in #2668
• Add DTensor compatible NS impl for Muon by @rwightman in #2669

New Contributors

• @Yuan-Jinghui made their first contribution in #2657
• @llukito made their first contribution in #2666

Full Changelog: v1.0.24...v1.0.25

Jan 5 & 6, 2025

• Patch Release 1.0.24 (fix for 1.0.23)
• Add new benchmark result csv files for inference timing on all models w/ RTX Pro 6000, 5090, and 4090 cards w/ PyTorch 2.9.1
• Fix moved module error in deprecated timm.models.layers import path that impacts legacy imports
• Release 1.0.23

Dec 30, 2025

• Add better NAdaMuon trained dpwee, dwee, dlittle (differential) ViTs with a small boost over previous runs
  • https://huggingface.co/timm/vit_dlittle_patch16_reg1_gap_256.sbb_nadamuon_in1k (83.24% top-1)
  • https://huggingface.co/timm/vit_dwee_patch16_reg1_gap_256.sbb_nadamuon_in1k (81.80% top-1)
  • https://huggingface.co/timm/vit_dpwee_patch16_reg1_gap_256.sbb_nadamuon_in1k (81.67% top-1)
• Add a ~21M param timm variant of the CSATv2 model at 512x512 & 640x640
  • https://huggingface.co/timm/csatv2_21m.sw_r640_in1k (83.13% top-1)
  • https://huggingface.co/timm/csatv2_21m.sw_r512_in1k (82.58% top-1)
• Factor non-persistent param init out of __init__ into a common method that can be externally called via init_non_persistent_buffers() after meta-device init.

Dec 12, 2025

• Add CSATV2 model (thanks https://github.com/gusdlf93) -- a lightweight but high res model with DCT stem & spatial attention. https://huggingface.co/Hyunil/CSATv2
• Add AdaMuon and NAdaMuon optimizer support to existing timm Muon impl. Appears more competitive vs AdamW with familiar hparams for image tasks.
• End of year PR cleanup, merge aspects of several long open PR
  • Merge differential attention (DiffAttention), add corresponding DiffParallelScalingBlock (for ViT), train some wee vits
    • https://huggingface.co/timm/vit_dwee_patch16_reg1_gap_256.sbb_in1k
    • https://huggingface.co/timm/vit_dpwee_patch16_reg1_gap_256.sbb_in1k
  • Add a few pooling modules, LsePlus and SimPool
  • Cleanup, optimize DropBlock2d (also add support to ByobNet based models)
• Bump unit tests to PyTorch 2.9.1 + Python 3.13 on upper end, lower still PyTorch 1.13 + Python 3.10

Dec 1, 2025

• Add lightweight task abstraction, add logits and feature distillation support to train script via new tasks.
• Remove old APEX AMP support

What's Changed

• Add val-interval argument by @t0278611 in #2606
• Add coord attn and some variants that I had lying around by @rwightman in #2617
• Distill fixups by @rwightman in #2598
• A simplification and some fixes for DropBlock2d. by @rwightman in #2620
• Other pooling... by @rwightman in #2621
• Experimenting with differential attention by @rwightman in #2314
• Differential + parallel attn by @rwightman in #2625
• AdaMuon impl w/ a few other ideas based on recent reading by @rwightman in #2626
• Csatv2 contribution by @rwightman in #2627
• Add HParams sections to hfdocs by @rwightman in #2630
• Upgrade GitHub Actions for Node 24 compatibility by @salmanmkc in #2633
• [BUG] Modify autocasting in fast normalization functions to handle optional weight params safely by @tesfaldet in #2631
• 'init_non_persistent_buffers' scheme by @rwightman in #2632
• Add docstrings to layer helper functions and modules by @raimbekovm in #2634
• refactor(scheduler): add type hints to CosineLRScheduler by @haru-256 in #2640
• A few misc weights to close out 2025 by @rwightman in #2639
• Update typing in other scheduler classes. Add unit tests. by @rwightman in #2641

New Contributors

• @t0278611 made their first contribution in #2606
• @salmanmkc made their first contribution in #2633
• @tesfaldet made their first contribution in #2631
• @raimbekovm made their first contribution in #2634
• @haru-256 made their first contribution in #2640

Full Changelog: v1.0.22...v1.0.24

Dec 30, 2025

• Add better NAdaMuon trained dpwee, dwee, dlittle (differential) ViTs with a small boost over previous runs
  • https://huggingface.co/timm/vit_dlittle_patch16_reg1_gap_256.sbb_nadamuon_in1k (83.24% top-1)
  • https://huggingface.co/timm/vit_dwee_patch16_reg1_gap_256.sbb_nadamuon_in1k (81.80% top-1)
  • https://huggingface.co/timm/vit_dpwee_patch16_reg1_gap_256.sbb_nadamuon_in1k (81.67% top-1)
• Add a ~21M param timm variant of the CSATv2 model at 512x512 & 640x640
  • https://huggingface.co/timm/csatv2_21m.sw_r640_in1k (83.13% top-1)
  • https://huggingface.co/timm/csatv2_21m.sw_r512_in1k (82.58% top-1)
• Factor non-persistent param init out of __init__ into a common method that can be externally called via init_non_persistent_buffers() after meta-device init.

Dec 12, 2025

• Add CSATV2 model (thanks https://github.com/gusdlf93) -- a lightweight but high res model with DCT stem & spatial attention. https://huggingface.co/Hyunil/CSATv2
• Add AdaMuon and NAdaMuon optimizer support to existing timm Muon impl. Appears more competitive vs AdamW with familiar hparams for image tasks.
• End of year PR cleanup, merge aspects of several long open PR
  • Merge differential attention (DiffAttention), add corresponding DiffParallelScalingBlock (for ViT), train some wee vits
    • https://huggingface.co/timm/vit_dwee_patch16_reg1_gap_256.sbb_in1k
    • https://huggingface.co/timm/vit_dpwee_patch16_reg1_gap_256.sbb_in1k
  • Add a few pooling modules, LsePlus and SimPool
  • Cleanup, optimize DropBlock2d (also add support to ByobNet based models)
• Bump unit tests to PyTorch 2.9.1 + Python 3.13 on upper end, lower still PyTorch 1.13 + Python 3.10

Dec 1, 2025

• Add lightweight task abstraction, add logits and feature distillation support to train script via new tasks.
• Remove old APEX AMP support

What's Changed

• Add val-interval argument by @t0278611 in #2606
• Add coord attn and some variants that I had lying around by @rwightman in #2617
• Distill fixups by @rwightman in #2598
• A simplification and some fixes for DropBlock2d. by @rwightman in #2620
• Other pooling... by @rwightman in #2621
• Experimenting with differential attention by @rwightman in #2314
• Differential + parallel attn by @rwightman in #2625
• AdaMuon impl w/ a few other ideas based on recent reading by @rwightman in #2626
• Csatv2 contribution by @rwightman in #2627
• Add HParams sections to hfdocs by @rwightman in #2630
• Upgrade GitHub Actions for Node 24 compatibility by @salmanmkc in #2633
• [BUG] Modify autocasting in fast normalization functions to handle optional weight params safely by @tesfaldet in #2631
• 'init_non_persistent_buffers' scheme by @rwightman in #2632
• Add docstrings to layer helper functions and modules by @raimbekovm in #2634
• refactor(scheduler): add type hints to CosineLRScheduler by @haru-256 in #2640
• A few misc weights to close out 2025 by @rwightman in #2639
• Update typing in other scheduler classes. Add unit tests. by @rwightman in #2641

New Contributors

• @t0278611 made their first contribution in #2606
• @salmanmkc made their first contribution in #2633
• @tesfaldet made their first contribution in #2631
• @raimbekovm made their first contribution in #2634
• @haru-256 made their first contribution in #2640

Full Changelog: v1.0.22...v1.0.23

Patch release for priority LayerScale initialization regression in 1.0.21

What's Changed

• Add some weights for efficientnet_x / efficientnet_h models by @rwightman in #2602
• Update result csvs by @rwightman in #2603
• Fix LayerScale ignoring init_values by @Ilya-Fradlin in #2605

New Contributors

• @Ilya-Fradlin made their first contribution in #2605

Full Changelog: v1.0.21...v1.0.22

Oct 16-20, 2025

• Add an impl of the Muon optimizer (based on https://github.com/KellerJordan/Muon) with customizations
  • extra flexibility and improved handling for conv weights and fallbacks for weight shapes not suited for orthogonalization
  • small speedup for NS iterations by reducing allocs and using fused (b)add(b)mm ops
  • by default uses AdamW (or NAdamW if nesterov=True) updates if muon not suitable for parameter shape (or excluded via param group flag)
  • like torch impl, select from several LR scale adjustment fns via adjust_lr_fn
  • select from several NS coefficient presets or specify your own via ns_coefficients
• First 2 steps of 'meta' device model initialization supported
  • Fix several ops that were breaking creation under 'meta' device context
  • Add device & dtype factory kwarg support to all models and modules (anything inherting from nn.Module) in timm
• License fields added to pretrained cfgs in code
• Release 1.0.21

What's Changed

• Add calculate_drop_path_rates helper by @rwightman in #2589
• Review huggingface_hub integration by @Wauplin in #2592
• Adding device/dtype factory_kwargs to modules and models by @rwightman in #2591
• Consistent license handling throughout timm by @alexanderdann in #2585
• Add impl of Muon optimizer. Fix #2580 by @rwightman in #2596
• Rename 'simple' flag for Muon to 'fallback' by @rwightman in #2599

New Contributors

• @alexanderdann made their first contribution in #2585

Full Changelog: v1.0.20...v1.0.21

Sept 21, 2025

• Remap DINOv3 ViT weight tags from lvd_1689m -> lvd1689m to match (same for sat_493m -> sat493m)
• Release 1.0.20

Sept 17, 2025

• DINOv3 (https://arxiv.org/abs/2508.10104) ConvNeXt and ViT models added. ConvNeXt models were mapped to existing timm model. ViT support done via the EVA base model w/ a new RotaryEmbeddingDinoV3 to match the DINOv3 specific RoPE impl
  • HuggingFace Hub: https://huggingface.co/collections/timm/timm-dinov3-68cb08bb0bee365973d52a4d
• MobileCLIP-2 (https://arxiv.org/abs/2508.20691) vision encoders. New MCI3/MCI4 FastViT variants added and weights mapped to existing FastViT and B, L/14 ViTs.
• MetaCLIP-2 Worldwide (https://arxiv.org/abs/2507.22062) ViT encoder weights added.
• SigLIP-2 (https://arxiv.org/abs/2502.14786) NaFlex ViT encoder weights added via timm NaFlexViT model.
• Misc fixes and contributions

What's Changed

• Pass init_values at hieradet_sam2 by @hassonofer in #2559
• Add mobileclip2 encoder weights by @rwightman in #2560
• Add support for Gemma 3n MobileNetV5 encoder weight loading by @rwightman in #2561
• Fix #2562, add siglip2 naflex vit encoder weights by @rwightman in #2564
• fix: create results_dir if missing before saving results by @zhima771 in #2576
• feat(validate): add precision, recall, and F1 metrics by @ha405 in #2568
• Allow user to ask for features other than image and label in ImageDataset by @grodino in #2571
• Add MobileCLIP2 image encoders by @rwightman in #2578
• Add DINOv3 support by @rwightman in #2579

New Contributors

• @hassonofer made their first contribution in #2559
• @zhima771 made their first contribution in #2576
• @ha405 made their first contribution in #2568

Full Changelog: v1.0.19...v1.0.20

Patch release for Python 3.9 compat break in 1.0.18

July 23, 2025

• Add set_input_size() method to EVA models, used by OpenCLIP 3.0.0 to allow resizing for timm based encoder models.
• Release 1.0.18, needed for PE-Core S & T models in OpenCLIP 3.0.0
• Fix small typing issue that broke Python 3.9 compat. 1.0.19 patch release.

July 21, 2025

• ROPE support added to NaFlexViT. All models covered by the EVA base (eva.py) including EVA, EVA02, Meta PE ViT, timm SBB ViT w/ ROPE, and Naver ROPE-ViT can be now loaded in NaFlexViT when use_naflex=True passed at model creation time
• More Meta PE ViT encoders added, including small/tiny variants, lang variants w/ tiling, and more spatial variants.
• PatchDropout fixed with NaFlexViT and also w/ EVA models (regression after adding Naver ROPE-ViT)
• Fix XY order with grid_indexing='xy', impacted non-square image use in 'xy' mode (only ROPE-ViT and PE impacted).

What's Changed

• Add ROPE support to NaFlexVit (axial and mixed), and support most (all?) EVA based vit models & weights by @rwightman in #2552
• Support set_input_size() in EVA models by @rwightman in #2554

Full Changelog: v1.0.17...v1.0.18

July 23, 2025

• Add set_input_size() method to EVA models, used by OpenCLIP 3.0.0 to allow resizing for timm based encoder models.
• Release 1.0.18, needed for PE-Core S & T models in OpenCLIP 3.0.0

July 21, 2025

• ROPE support added to NaFlexViT. All models covered by the EVA base (eva.py) including EVA, EVA02, Meta PE ViT, timm SBB ViT w/ ROPE, and Naver ROPE-ViT can be now loaded in NaFlexViT when use_naflex=True passed at model creation time
• More Meta PE ViT encoders added, including small/tiny variants, lang variants w/ tiling, and more spatial variants.
• PatchDropout fixed with NaFlexViT and also w/ EVA models (regression after adding Naver ROPE-ViT)
• Fix XY order with grid_indexing='xy', impacted non-square image use in 'xy' mode (only ROPE-ViT and PE impacted).

What's Changed

• Add ROPE support to NaFlexVit (axial and mixed), and support most (all?) EVA based vit models & weights by @rwightman in #2552
• Support set_input_size() in EVA models by @rwightman in #2554

Full Changelog: v1.0.17...v1.0.18

July 7, 2025

• MobileNet-v5 backbone tweaks for improved Google Gemma 3n behaviour (to pair with updated official weights)
  • Add stem bias (zero'd in updated weights, compat break with old weights)
  • GELU -> GELU (tanh approx). A minor change to be closer to JAX
• Add two arguments to layer-decay support, a min scale clamp and 'no optimization' scale threshold
• Add 'Fp32' LayerNorm, RMSNorm, SimpleNorm variants that can be enabled to force computation of norm in float32
• Some typing, argument cleanup for norm, norm+act layers done with above
• Support Naver ROPE-ViT (https://github.com/naver-ai/rope-vit) in eva.py, add RotaryEmbeddingMixed module for mixed mode, weights on HuggingFace Hub

• Some cleanup of ROPE modules, helpers, and FX tracing leaf registration
• Preparing version 1.0.17 release

What's Changed

• Adding Naver rope-vit compatibility to EVA ViT by @rwightman in #2529
• Update no_grad usage to inference_mode if possible by @GuillaumeErhard in #2534
• Add a min layer-decay scale clamp, and no optimization threshold to exclude groups from optimization by @rwightman in #2537
• Add stem_bias option to MNV5. Resolve the norm layer so can pass string. by @rwightman in #2538
• Add flag to enable float32 computation for normalization (norm + affine) by @rwightman in #2536
• fix: mnv5 conv_stem bias and GELU with approximate=tanh by @RyanMullins in #2533
• Fixup casting issues for weights/bias in fp32 norm layers by @rwightman in #2539
• Fix H, W ordering for xy indexing in ROPE by @rwightman in #2541
• Fix 3 typos in README.md by @robin-ede in #2544

New Contributors

• @GuillaumeErhard made their first contribution in #2534
• @RyanMullins made their first contribution in #2533
• @robin-ede made their first contribution in #2544

Full Changelog: v1.0.16...v1.0.17