Last updated: November 16, 2026
Hugging Face Transformers 5.23, released November 13, 2026, adds a single exporters API: every export to Dynamo, ONNX Runtime, OpenVINO or ExecuTorch now returns an ExportedModel you can call, push to the Hub and reload. It also adds Inkling 2 Small, LFM3, LFM3-VL and other models, plus several breaking changes.
Contents
- What changed in Transformers 5.23
- The new exporters API
- How to export a model to ONNX
- New model support
- Breaking changes to check
- Transformers 5.23 compared with 5.22
- How to install or upgrade
- Limitations
- Frequently asked questions
What changed in Transformers 5.23
Transformers is the Hugging Face Python library for loading, running, training and now exporting open models. Version 5.23.0 was published to GitHub and PyPI on November 13, 2026. According to the 5.23.0 release notes, the main changes are:
| Area | Change in 5.23.0 |
|---|---|
| Exporters | One API for Dynamo, ONNX Runtime, OpenVINO and ExecuTorch that returns a callable, shareable ExportedModel |
| ONNX Runtime | Optional fused transformer ops at export time with ort_fusions |
| OpenVINO | Fused CPU kernels for attention, RoPE, MoE experts and int8/int4 weights |
| New models | Inkling2, JEV, LFM3, Dots 3 Note Preview, LFM3-VL, IQuestCoder, InternLM3 |
| Parallelism | Selectable expert-parallel implementation: all-reduce for inference, all-to-all for training |
| Fixes | Vision, cache, attention, generation, quantization and audio corrections |
The new exporters API
Before 5.23, each export path had its own return type. Now every export returns an ExportedModel. You call it like the original model, with the same keyword arguments and the same output class. The exporters documentation lists four exporters:
DynamoExportersavesmodel.pt2for any PyTorch runtime or ahead-of-time compilation.OnnxExportersavesmodel.onnxfor ONNX Runtime, TensorRT or OpenVINO.ExecutorchExportersavesmodel.ptefor mobile and edge devices.OpenVINOExportersavesopenvino_model.xmland.binfor Intel CPU, GPU and NPU.
An export saves and loads like a model, optionally in a subfolder of the model repo. Its export.json file records the config, input shapes, package versions and input and output names. Quantized exports made with PT2E, NNCF or ONNX Runtime save and run the same way.
Hugging Face reports large speedups in its own tests. On EmbeddingGemma2 with 8 sequences of 128 tokens, ONNX Runtime with the TensorRT execution provider ran 7.6 times faster than eager PyTorch and 1.4 times faster than torch.compile on an A100. OpenVINO ran 4.1 times faster than eager on a Xeon Platinum 8570. OpenVINO decode is 2.4 to 3.9 times faster than in earlier exports. Treat these as vendor figures for one setup, not a guarantee for your model.
How to export a model to ONNX
The release notes give this workflow for an ONNX export:
- Upgrade with
pip install -U transformersand install ONNX Runtime for your target device. - Load the model and tokenizer with
AutoModel.from_pretrainedandAutoTokenizer.from_pretrained, then tokenize sample inputs withreturn_tensors="pt". - Import
OnnxExporter,OnnxConfigandExportedModelfromtransformers.exporters. - Run
OnnxExporter().export(model, inputs, OnnxConfig(dynamic=True, ort_fusions="cuda")). Use"cpu","webgpu"or"dml"for other targets. - Save the result with
save_pretrained, or share it withpush_to_hub("your-org/model-onnx"). - Reload it with
ExportedModel.from_pretrained("your-org/model-onnx", device="cuda")and call it with the same inputs as the original model.
One security point from the documentation: a Dynamo .pt2 file is a pickle, so from_pretrained refuses to load it unless you pass trust_remote_code=True. Only do that for exports you trust. ONNX, OpenVINO and ExecuTorch programs load without it. A directory holds one export, so keep several formats in separate subfolders such as onnx/ and openvino/.
New model support
Seven model families were added in 5.23.0:
- Inkling2: Thinking Machines Lab’s Inkling 2 Small, a text and image model with 594B total and 17B active parameters. See our Inkling 2 Small open weights guide.
- LFM3 and LFM3-VL: Liquid AI’s third-generation hybrid models for text and for images and video. Our LFM3 guide covers the family.
- JEV: an experimental decision head that scores single-token labels. It also reads the 255 answer codes of pplx-decider, covered in our Perplexity Decider v1.3 guide, and runs through an experimental
decisionpipeline. - Dots 3 Note Preview: a mixture-of-experts model that takes text, image, video and audio input, with BF16 and FP8 inference.
- IQuestCoder and InternLM3: both load with the Llama classes. Looped IQuest-Coder-V1 checkpoints are not supported.
Breaking changes to check
The release lists 12 breaking changes. The ones most likely to affect users are:
export()returns anExportedModel, an ONNX program is anonnx_ir.Model, and theoutput_pathoptions are gone in favor ofsave_pretrained.- The
extra now installs only torchaudio, soundfile and soxr. librosa and phonemizer were dropped. Wav2Vec2ProcessorWithLMand pyctcdecode/kenlm decoding are deprecated.- The
ep_routerexpert-parallel style was removed, and MXFP4 GPT-OSS experts run throughep_dispatch_experts. - Rope deltas were dropped from
ModelOutput, and processing defaults moved to the class. - The serve command changed its attention setup for Mac users.
Transformers 5.23 compared with 5.22
Version 5.22.0, released October 29, centered on one model. The 5.22.0 release notes describe these changes:
| Release | Main additions | Breaking changes |
|---|---|---|
| 5.22.0 (Oct 29) | Mistral Large 4 support; LoRA and other adapter training on FP8 weights, with MoE backward peak memory cut from about 28.9 GiB to 2.8 GiB at 8k tokens | 1 (expert-parallel logic moved to its own file) |
| 5.23.0 (Nov 13) | Unified exporters, seven new model families, selectable expert-parallel implementation | 12 |
If you serve models rather than export them, a dedicated engine may still fit better. Our guides to vLLM 0.32 and llama.cpp 1.0.0 cover those options.
How to install or upgrade
The installation guide says Transformers is tested on Python 3.11+ and PyTorch 2.6+. It recommends a uv virtual environment:
- Create and activate an environment with
uv venv .envandsource .env/bin/activate. - For NVIDIA GPUs, run
uv pip install "transformers[torch]". For CPU only, install PyTorch from the CPU wheel index first, thenuv pip install transformers. - Test the install with a short pipeline call, as shown in the guide.
Limitations
- Experimental exporters. Hugging Face says the exporters are experimental and patch around upstream bugs, so pin versions in production.
- Experimental JEV. The JEV head and decision pipeline may change and should be used with caution in production.
- Upgrade work. Code that relied on old export return types,
output_pathor the old audio extra needs changes.
Frequently asked questions
What is new in Hugging Face Transformers 5.23?
Transformers 5.23.0, released November 13, 2026, adds a unified exporters API that returns an ExportedModel for Dynamo, ONNX Runtime, OpenVINO and ExecuTorch, seven new model families including Inkling 2 Small, LFM3 and LFM3-VL, an experimental JEV decision head, and a dozen breaking changes.
How do I export a Transformers model to ONNX in 5.23?
Create an OnnxExporter, call export with the model, sample inputs and an OnnxConfig, then save the result with save_pretrained or push_to_hub. Load it later with ExportedModel.from_pretrained and call it with the same keyword arguments as the original model.
Is the Transformers exporters API stable?
No. The documentation calls the exporters experimental and says many patches work around specific upstream bugs. Hugging Face advises pinning the Torch, ONNX Script, ONNX Runtime and ExecuTorch versions used in its test suite for production tooling.
Which Python and PyTorch versions does Transformers 5.23 need?
PyPI lists Python 3.11 or newer for 5.23.0. The installation guide says Transformers has been tested on Python 3.11+ and PyTorch 2.6+.
Will my old export code break after upgrading?
It may. In 5.23 export returns an ExportedModel instead of a raw program, an ONNX program is an onnx_ir.Model, and the output_path options on OnnxConfig and OpenVINOConfig were removed in favor of save_pretrained.
How do I upgrade to Transformers 5.23?
Run pip install -U transformers, or uv pip install -U transformers in a uv environment. Then check the version with python -c "import transformers; print(transformers.__version__)".
Sources
- Transformers v5.23.0 release notes, GitHub (November 13, 2026)
- Transformers v5.22.0 release notes, GitHub (October 29, 2026)
- Exporters, Transformers documentation
- Installation, Transformers documentation
Featured image: Hugging Face logo, public domain, via Wikimedia Commons. Hugging Face is a trademark of Hugging Face, Inc.