Overview

🤗 Optimum handles the export of PyTorch models to ONNX in the exporters.onnx module. It provides classes, functions, and a command line interface to perform the export easily.

Supported architectures from 🤗 Transformers:

• Arcee
• AST
• Audio Spectrogram Transformer
• Albert
• Bart
• Beit
• Bert
• BlenderBot
• BlenderBotSmall
• Bloom
• Camembert
• ChineseCLIP
• CLIP
• CodeGen
• Cohere
• ConvBert
• ConvNext
• ConvNextV2
• D-FINE
• Data2VecAudio
• Data2VecText
• Data2VecVision
• Deberta
• Deberta-v2
• Decision Transformer
• DeepSeek-V3
• Deit
• Detr
• DINOv2
• DistilBert
• Donut-Swin
• Electra
• Encoder Decoder
• ESM
• Falcon
• Flaubert
• Gemma
• Gemma 2
• GLM
• GPT-2
• GPT-BigCode
• GPT-J
• GPT-Neo
• GPT-NeoX
• OPT
• Granite
• GroupVit
• Helium
• Hiera
• Hubert
• IBert
• InternLM2
• LayoutLM
• LayoutLM-v3
• Lilt
• Levit
• LongT5
• Llama
• M2-M100
• Marian
• MarkupLM
• MaskFormer
• MBart
• MetaClip2
• MGP-STR
• Mistral
• MobileBert
• MobileVit
• MobileNet v1
• MobileNet v2
• ModernBert
• MPNet
• MT5
• Musicgen (text-conditional only)
• Nemotron
• Nystromformer
• OLMo
• OLMo2
• OWL-ViT
• PatchTST
• PatchTSMixer
• Pegasus
• Perceiver
• Phi
• Phi3
• Pix2Struct
• PoolFormer
• PVT
• Qwen2(Qwen1.5)
• Qwen3
• Qwen3-MoE
• RegNet
• RemBERT
• ResNet
• Roberta
• Roformer
• RT-DETR
• RT-DETRv2
• SAM
• Segformer
• SEW
• SEW-D
• Speech2Text
• SigLIP
• SmolLM3
• SpeechT5
• Splinter
• SqueezeBert
• StableLM
• Swin
• SwinV2
• T5
• Table Transformer
• TROCR
• UniSpeech
• UniSpeech SAT
• Vision Encoder Decoder
• VisualBert
• Vit
• VitMAE
• VitMSN
• Wav2Vec2
• Wav2Vec2 Conformer
• WavLM
• Whisper
• XLM
• XLM-Roberta
• Yolos

Supported architectures from 🤗 Diffusers:

• Stable Diffusion

Supported architectures from 🤗 Timm:

• Adversarial Inception v3
• AdvProp (EfficientNet)
• Big Transfer (BiT)
• CSP-DarkNet
• CSP-ResNet
• CSP-ResNeXt
• DenseNet
• Deep Layer Aggregation
• Dual Path Network (DPN)
• ECA-ResNet
• EfficientNet
• EfficientNet (Knapsack Pruned)
• Ensemble Adversarial Inception ResNet v2
• ESE-VoVNet (Partial support with static shapes)
• FBNet
• (Gluon) Inception v3
• (Gluon) ResNet
• (Gluon) ResNeXt
• (Gluon) SENet
• (Gluon) SE-ResNeXt
• (Gluon) Xception
• HRNet
• Instagram ResNeXt WSL
• Inception ResNet v2
• Inception v3
• Inception v4
• (Legacy) SE-ResNet
• (Legacy) SE-ResNeXt
• (Legacy) SENet
• MixNet
• MnasNet
• MobileNet v2
• MobileNet v3
• NASNet
• Noisy Student (EfficientNet)
• PNASNet
• RegNetX
• RegNetY
• Res2Net
• Res2NeXt
• ResNeSt
• ResNet
• ResNet-D
• ResNeXt
• RexNet
• SE-ResNet
• SelecSLS
• SE-ResNeXt
• SK-ResNet
• SK-ResNeXt
• SPNASNet
• SSL ResNet
• SWSL ResNet
• SWSL ResNeXt
• TResNet
• Wide ResNet
• Xception

Supported architectures from Sentence Transformers:

• All Transformer and CLIP-based models.