ZeroModels Releases 118 Keras 3 Model Families Across Three Backends

ZeroModels has released with 118 pretrained model families implemented in Keras 3, giving developers one model interface across JAX, PyTorch, and TensorFlow. The new library’s central promise is portability: it says a project can change its Keras backend without rewriting model code.

The catalog spans vision, language, multimodal, and speech tasks. Listed families include DETR for object detection, SegFormer for segmentation, Depth Anything V2 for depth estimation, SAM 3 for promptable masks, Whisper for speech recognition, and language-model families including Qwen3, Gemma 4, GLM-5, and GPT-OSS.

ZeroModels describes its implementations as pure Keras 3 and says neither Transformers nor PyTorch is required at runtime. That separates the library from the original runtime stacks used by many of the model families it converts, while leaving the same Keras-facing code available across its three supported backends.

One loader, several routes to weights

Its main entry point is called from_weights. The API can load a preconverted Keras repository, convert an upstream checkpoint from a bare variant name, or load a compatible Hugging Face repository through an hf: prefix. It can also construct an architecture without loading weights, which gives developers a route to start from a randomly initialized model.

Preconverted weights are hosted through the zeromodels organization on the Hugging Face Hub. The documentation uses the same identifier for a model and its processor, and says that pairing keeps a processor’s resolution and normalization settings aligned with the checkpoint it loads.

Portability does not erase hardware limits

The library also offers weight-only int8, int4, fp8, and MXFP4 quantization through the same loading call. For GPT-OSS 120B, ZeroModels says its mixture-of-experts weights can remain packed in MXFP4 and be dequantized during use, keeping the model near 66 GB instead of roughly 130 GB for an fp32 expansion. Those are project-provided memory figures, not an independent measure of deployment requirements.