Models¶
Every model here is a pure-Keras 3 port with weights converted from the original checkpoints. Pages are grouped by modality and listed alphabetically, and each one follows the same shape: API, preprocessing, variants, a runnable example with its measured output, data format, and how to load fine-tuned weights.
The shared machinery is documented separately, in Main Classes and Configuration.
Text models¶
Encoders for embeddings and masked language modelling, and decoder LLMs, dense and mixture-of-experts.
- BERT
- Cohere
- Cohere 2
- Cohere 2 MoE
- DeBERTa
- DeepSeek-V2
- DeepSeek-V3
- DeepSeek-V4
- Gemma
- Gemma 2
- GLM
- GLM-4
- GLM-4 MoE
- GLM-5 MoE
- GPT
- GPT-2
- GPT-OSS
- Llama
- Llama 2
- Llama 4
- MiniMax
- MiniMax M2
- Mistral
- Mixtral
- Qwen
- Qwen2
- Qwen2 MoE
- Qwen3
- Qwen3 MoE
- Qwen3.5
- Qwen3.5 MoE
- RoBERTa
- XLM-RoBERTa
Vision models¶
Detection, segmentation, monocular depth, and self-supervised backbones.
- Classification backbones
- D-FINE
- DeepLabV3
- Depth Anything V1
- Depth Anything V2
- DETR
- DINO
- DINOv2
- DINOv3
- EoMT
- Mask2Former
- MaskFormer
- MobileViT
- MobileViTV2
- RF-DETR
- RT-DETR
- RT-DETRv2
- SAM
- SAM 2
- SegFormer
Audio models¶
Speech recognition, and speech-aware language models that take audio and text together.
Multimodal models¶
Vision-language encoders, generative VLMs, and grounding across detection, OCR, pointing and referring.