Models

Metalhead has 335 ported architectures in 25 families. Build one with create_model("<architecture>"), or load a Hub checkpoint of it with create_model("<organisation>/<repository>"; pretrained = true).

Models are grouped by library and family, a model line. A timm family is usually one of timm's model modules: Wide ResNet, ResNeXt and SE-ResNet belong to the ResNet family; Res2Net and aligned Xception have their own families; the three Inception modules share one. A transformers or diffusers family, such as Granite or FLUX.2, can span several modules.

Browse the same families from Julia with list_models:

families = list_models(; byfamily = true)
families["resnet"]
list_models(; byfamily = true, pretrained = true)["resnet"]

The 94 checkpoints listed here were compared with PyTorch's outputs (see How models are validated). Bare architecture names load the first one with pretrained = true. Other checkpoints of the same architectures load with the same strict checks. A dash means no checkpoint has been validated for that architecture.

• Models
• • Vision models (timm)
  • • Aligned Xception
    • ConvMixer
    • ConvNeXt
    • DenseNet
    • EfficientNet, MobileNetV1/V2, MNASNet and MixNet
    • Inception-v3, Inception-v4 and Inception-ResNet-v2
    • MLP-Mixer, ResMLP and gMLP
    • MobileNetV3/V4, FBNetV3 and LCNet
    • Res2Net
    • ResNet
    • VGG
    • Vision Transformer (ViT)
  • Language, text encoder, vision-language, image-text and segmentation models (transformers)
  • • BERT
    • CLIP
    • Gemma 3
    • Granite
    • Llama
    • ModernBERT
    • Qwen
    • SAM 3
    • SigLIP and SigLIP 2
    • XLM-RoBERTa
  • Image generation (diffusers)
  • • FLUX.2
    • Z-Image
  • Text embeddings (sentence-transformers)
  • • Sentence Transformers

Vision models (timm)

Aligned Xception

xception_aligned: 5 architectures. Source: xception.jl.

ConvMixer

convmixer: 3 architectures. Source: convmixer.jl.

ConvNeXt

convnext: 8 architectures. Source: convnext.jl.

DenseNet

densenet: 6 architectures. Source: densenet.jl.

EfficientNet, MobileNetV1/V2, MNASNet and MixNet

efficientnet: 99 architectures. Source: efficientnet.jl.

Inception-v3, Inception-v4 and Inception-ResNet-v2

inception: 3 architectures. Source: inception_resnet_v2.jl, inception_v3.jl, inception_v4.jl.

MLP-Mixer, ResMLP and gMLP

mlp_mixer: 15 architectures. Source: mlp_mixer.jl.

MobileNetV3/V4, FBNetV3 and LCNet

mobilenetv3: 33 architectures. Source: mobilenetv3.jl.

Res2Net

res2net: 9 architectures. Source: res2net.jl.

ResNet

resnet: 58 architectures. Source: resnet.jl.

VGG

vgg: 8 architectures. Source: vgg.jl.

Vision Transformer (ViT)

vision_transformer: 51 architectures. Source: vit.jl.

Language, text encoder, vision-language, image-text and segmentation models (transformers)

BERT

bert: 2 architectures. Source: bert.jl.

create_model builds the sentence-transformers model of sentence-transformers/all-MiniLM-L6-v2, BAAI/bge-small-en-v1.5 and the other repositories with a modules.json, whose BertModel the references validate. A BertModel checkpoint without pooler weights needs add_pooling_layer = false.

CLIP

clip: 5 architectures. Source: clip.jl.

The main branch of openai/clip-vit-base-patch32 has no safetensors weights: its validated commit is the pull request of Hugging Face's conversion bot that adds them, which create_model loads by default.

Gemma 3

gemma3: 1 architecture. Source: gemma3.jl.

Granite

granite: 4 architectures. Source: granite.jl, granitemoe.jl.

Granite 4.0 models load without Mamba layers, so the hybrid "H" models (granite-4.0-h-*) don't load.

Llama

llama: 1 architecture. Source: llama.jl.

ModernBERT

modernbert: 3 architectures. Source: modernbert.jl.

Qwen

qwen: 8 architectures. Source: qwen2.jl, qwen3.jl, qwen3_moe.jl, qwen3_5.jl, qwen3_5_moe.jl, qwen3_5_vision.jl.

Qwen3-30B-A3B and Qwen3.6-35B-A3B were validated on their first 4 decoder layers, with their embeddings, head and, for Qwen3.6, vision encoder; Qwen/Qwen3.8-27B with its first 8 decoder layers (see How models are validated).

SAM 3

sam3: 2 architectures. Source: sam3.jl.

jetjodh/sam3 is an ungated copy of the weights of Meta's gated facebook/sam3. Only the detector of Sam3VideoModel is ported, not its tracker.

SigLIP and SigLIP 2

siglip: 2 architectures. Source: siglip2.jl, siglip.jl.

The NaFlex checkpoints of SigLIP 2 (google/siglip2-*-naflex) are Siglip2Models, which take images of any aspect ratio; the others, of fixed resolution, are SiglipModels.

XLM-RoBERTa

xlm_roberta: 2 architectures. Source: xlm_roberta.jl.

BAAI/bge-m3 stores its weights in pytorch_model.bin only, which Metalhead doesn't read; BAAI/bge-reranker-v2-m3 has the same architecture and is validated.

Image generation (diffusers)

FLUX.2

flux2: 3 architectures. Source: autoencoder_kl_flux2.jl, pipeline_flux2_klein.jl, transformer_flux2.jl.

Z-Image

z_image: 3 architectures. Source: autoencoder_kl.jl, pipeline_z_image.jl, transformer_z_image.jl.

AutoencoderKL is diffusers' autoencoder of Stable Diffusion and FLUX.1 too, validated with the one of Z-Image.

Text embeddings (sentence-transformers)

Sentence Transformers

sentence_transformers: 1 architecture. Source: model.jl.

A repository with a modules.json loads as a SentenceTransformer, whose Transformer module holds a transformers model of a validated model_type (qwen3).