Hugging Face’s Transformers v5.23.0 adds support for several AI model families and gives exported models a more consistent way to move between runtimes. The release also reports substantial inference speed-ups on specific hardware, making this a practical upgrade for teams deploying models, not just a long list of merged code.
Hugging Face Watch analysis
What happened
The new release adds support for models including Inkling2, Liquid AI’s LFM3, Dots 3 Note Preview, IQuestCoder and InternLM3. It also introduces JEV, an experimental decision head whose APIs and behaviour may change.
The biggest infrastructure change is a unified export-and-load workflow. Exported models can be saved to the Hugging Face Hub and loaded on the backend’s own runtime, with support spanning Dynamo, ONNX Runtime, OpenVINO and ExecuTorch. The release notes report OpenVINO decoding 2.4–3.9 times faster than before on the specified workloads. They also report ONNX Runtime with TensorRT running EmbeddingGemma2 7.6 times faster than eager PyTorch on an A100.
Why it matters
A model that works in one framework but is awkward to export, share or run elsewhere is only part-way to being useful. A common save-and-load path could make it easier for teams to move models between runtimes, while the reported speed gains offer a reason for operators to test the updated backends on their own workloads.
The release also includes fixes for vision-model behaviour, distributed training and inference, and cache handling. Those are less headline-friendly than a new model, but fewer broken edge cases can matter more to a production team than another architecture name to memorise.
Our read
This is a substantial release with a clear practical centre: model support on one side, a tidier route from export to deployment on the other. The performance figures are tied to particular models and hardware, so they are a useful prompt to benchmark, not a promise that every workload will suddenly sprint.
What to watch
- Whether the unified exporter workflow makes it easier to share and run models across backends.
- How the reported speed-ups hold up on other models and hardware.
- How the experimental JEV decision pipeline changes as its API develops.
Discussion spark: When moving a model into production, what matters more to you: broader model support or a faster, more consistent export path?
Sources and evidence
- Release v5.23.0 (13 November 2026, 14:10 UTC)
not affiliated with or endorsed by Hugging Face