OpenToolslogo
ToolsExpertsNewsletter
AdvertiseLearn AI
  1. Home
  2. /
  3. News
  4. /
  5. Hugging Face Transformers 5.23.0 adds seven model integrations, rewrites the exporter API and lists twelve breaking changes
Updated 1 minute ago

Share this article

PostShare

In This Article

  • What v5.23.0 is, and what it is not
  • Seven model integrations, at different maturity levels
  • The exporter rewrite is the change most likely to touch deployment tooling
  • Twelve breaking changes, most stated as titles only
  • Inference, loading and training fixes that affect upgrades
  • What the official notes do not establish

AI news in your inbox

Weekly updates on tools, models, and the companies building them.

Subscribe

Footer

Company name

The right AI tool is out there. We'll help you find it.

LinkedInX

Knowledge Hub

  • News
  • Resources
  • Newsletter
  • Blog
  • AI Tool Reviews
  • YouTube Summary
  • YouTube Transcript Generator

Industry Hub

  • AI Companies
  • AI Tools
  • AI Models
  • MCP Servers
  • Muse Connectors
  • AI Tool Categories
  • Top AI Use Cases

For Builders

  • Submit a Tool
  • Submit a MCP
  • Experts & Agencies
  • Advertise
  • Compare Tools
  • Favourites

Legal

  • Privacy Policy
  • Terms of Service

© 2026 OpenTools - All rights reserved.

Hugging Face Transformers 5.23.0 adds seven model integrations, rewrites the exporter API and lists twelve breaking changes
Source: Hugging Face website homepage viewed through a magnifying glass. Archival photograph taken 25 August 2023 (12:29, Canon EOS 5D Mark IV) by Jernej Furman from Slovenia, originally posted to Flickr by focusonmore.com (https://flickr.com/photos/91261194@N06/53146954891), transferred to Wikimedia Commons. This is a 2023 image of the Hugging Face platform's website, not a depiction of the Transformers 5.23.0 release. File as published here was resized proportionally for web publication (6,720 × 4,480 → 2,400 × 1,600) with embedded metadata removed; no crop, overlay, compositing or AI generation was applied. Credit: Jernej Furman / focusonmore.com, licensed under CC BY 2.0 (https://creativecommons.org/licenses/by/2.0); changes of resizing and metadata removal indicated per the license. Image: Jernej Furman (Slovenia), posted to Flickr by focusonmore.com. CC BY 2.0. · Source image · CC BY 2.0

Hugging Face Transformers 5.23.0 adds seven model integrations, rewrites the exporter API and lists twelve breaking changes

By OpenTools Team · November 18, 2026

Hugging Face published Transformers v5.23.0 on 13 November 2026. The release adds seven model integrations — Inkling 2 Small, JEV, LFM3, LFM3‑VL, Dots 3 Note Preview, IQuestCoder and InternLM3 — rebuilds the exporter API around an ExportedModel object, and flags twelve breaking changes. For deployers the practical pinch points are the exporter API break, the slimmer `[audio]` extra that drops librosa and phonemizer, and the deprecation of Wav2Vec2ProcessorWithLM LM decoding. For fine‑tuners, expert parallelism gains a selectable implementation and the Trainer reduces FSDP/HSDP gradients only on the last micro‑batch. Most breaking entries are one‑line titles that do not state code impact, and the notes give no consolidated 5.23 upgrade checklist and no numeric minimum for the `kernels` dependency. The exporter speedups are the project's own figures, with methodology unstated and no independent verification.

What v5.23.0 is, and what it is not

Hugging Face published Release v5.23.0 of its Transformers library on 13 November 2026, and the release record returned by the GitHub releases API lists it as a normal release rather than a prerelease, carrying the tag v5.23.0.1

This is the open-source Python library for loading, running and training models, not the film franchise — and the date matters: the release landed on 13 November, so it is not a same-day story.

OpenTools has covered this library before, most recently the Transformers 5.21 MiniCPM-OCR integration. That article addressed a different minor version and a document-recognition workflow, so it is background rather than a preview of what 5.23.0 changes.

Seven model integrations, at different maturity levels

The release notes add seven model integrations:1

  • Inkling 2 Small — a multimodal model from Thinking Machines Lab that accepts text and image inputs and generates text. Its language backbone has 48 mixture-of-experts decoder layers, each routing tokens to 8 of 384 experts, for 594B total and 17B active parameters; a Perception Encoder-based vision transformer supplies one embedding per 40x40 pixel patch.
  • LFM3 — Liquid AI's third-generation hybrid model, combining gated short convolutions with grouped-query attention split between full-attention and sliding-window layers.
  • LFM3-VL — a vision-language family built on LFM3 that takes images and video at native resolution plus text, with a projector merging each 2x2 block of patch features into one language token.
  • Dots 3 Note Preview — a mixture-of-experts causal language model that accepts native text, image, video and audio inputs, using a shared vision encoder and a Whisper-style audio encoder, with BF16 and FP8 inference supported.
  • IQuestCoder — code models from IQuest Lab in 7B, 14B and 40B sizes with base, instruct and thinking variants. Its architecture matches Llama, so the auto classes load it through LlamaModel and LlamaForCausalLM with only IQuestCoderConfig specific. The looped IQuest-Coder-V1-*-Loop-* checkpoints use a different architecture and are not supported.
  • InternLM3 — an 8B instruction-tuned model from Shanghai AI Laboratory that likewise loads through the Llama implementation, with InternLM3Config requiring its attention_bias and mlp_bias values to be equal, as they are in the released checkpoints.
  • JEV — a decision head that scores single-token labels with a linear layer at the last token of the prompt, reading head files shipped beside the weights and able to run through the decision pipeline or decide.

JEV is the one addition the notes explicitly mark as unsettled: they describe it and the decision pipeline as currently experimental, say their APIs, configuration and behaviour may change, and warn they should be "used with caution in production environments".1 If you plan to put a decision head into production, that note is the relevant boundary rather than the model's inclusion in the library.

The release notes link documentation pages for each addition, but those pages were not part of the evidence reviewed here, so checkpoint availability, licensing and hardware requirements for the new models are not established by this article.

The exporter rewrite is the change most likely to touch deployment tooling

Exporters were rebuilt around a single object. Every export now returns an ExportedModel you call the way you call the original model, push to the Hub, and load back on the backend's own runtime, with the same API spanning Dynamo, ONNX Runtime, OpenVINO and ExecuTorch.1 Quantized exports (PT2E, NNCF, ONNX Runtime) save and run the same way.

This is also a breaking API change, and one of the few the notes spell out: export() returns an ExportedModel, an ONNX program is an onnx_ir.Model, the output_path options on OnnxConfig and OpenVINOConfig are gone in favour of save_pretrained, and custom backends register with register_backend. Any script that passes output_path or expects the old return type will need editing.

The notes also carry vendor-stated speedups, which are the project's own figures rather than independent measurements:

  • OpenVINO exports now reach OpenVINO's fused CPU kernels, which the notes say makes decode 2.4–3.9x faster than before.
  • On EmbeddingGemma2 (8 x 128 tokens), ONNX Runtime with the TensorRT EP is reported 7.6x faster than eager PyTorch and 1.4x faster than torch.compile on an A100.
  • OpenVINO is reported 4.1x faster than eager and 2.2x faster than torch.compile on a Xeon Platinum 8570 (4.9x and 2.6x with int8 weights).

The notes do not state the batch size, precision or baseline version behind the "2.4–3.9x faster than before" comparison, so treat that range as a claim about the exporter path rather than a prediction for your workload.

Twelve breaking changes, most stated as titles only

The notes open a dedicated breaking-changes section with twelve entries.1 For most of them the entry is a one-line title, so the notes establish that a change happened without stating what code it breaks or what replaces it.

The entries that come with usable detail are the two audio changes and the exporter rewrite:

  • The [audio] extra is slimmed to torchaudio, soundfile and soxr, dropping librosa and phonemizer — so the new install surface is stated, even if which pipelines break beyond the dependency change is not.
  • Wav2Vec2ProcessorWithLM and the pyctcdecode/kenlm language-model decoding path are deprecated, with no removal date given.
  • The exporter API break described above.

The remaining entries are titles without stated impact: removing the ep_router expert-parallel style, running MXFP4 GPT-OSS experts through ep_dispatch_experts, adding a deepgemm_megamoe expert-parallel implementation, moving processing defaults to the class, dropping rope deltas from ModelOutput, the kernels minimum-version bump, an attention change for Mac users in serve, "safe-izing" pi0 graphs, and building the tp/ep_plan/fsdp_plan once.

Practically, that means the notes tell you which areas to inspect — expert parallelism, processor defaults, rope deltas, kernels, Mac serving — but not what your own code must change. The linked pull requests are where the detail sits.

Inference, loading and training fixes that affect upgrades

Several fixes land on paths that deployments commonly hit:1

  • Cache correctness: Gemma4 static cache attending to empty slots in KV-shared layers at prefill, Qwen VL positions when a text cache is reused with a new image, and DeepSeek-V4 compression cache reordering during beam search.
  • Generation: grouped-query attention decoding now folds query heads onto their KV head instead of repeating K and V, which the notes describe as avoiding a copy of the KV cache; padded generation, batched decoding and pipeline metadata each get a fix.
  • Loading: mismatched keys are no longer silently ignored.
  • Quantization: FP8 modules_to_not_convert entries are renamed like weight names, so modules such as f_a_proj and f_b_proj are no longer wrongly quantized, and MXFP4 experts dequantize to the requested dtype instead of always using bfloat16.

For fine-tuning and distributed runs, the notes describe a selectable ep_implementation for expert parallelism — all-reduce for CUDA-graph-friendly inference, all-to-all for training — a Trainer change that reduces FSDP/HSDP gradients only on the last micro-batch, and token-count normalization corrected to use true division. The release also raises the minimum kernels version and bumps mamba-ssm kernels, and the project's Docker and CI images move to torch 2.15 / cu132. The notes do not say whether torch 2.15 is a user-facing minimum or only the environment the project tests against.

What the official notes do not establish

Three limits are worth carrying into an upgrade decision.

First, there is no consolidated migration checklist. The notes list breaking changes, deprecated audio components and the exporter API break, and then point at pull requests; a version-5 migration guide is referenced for one vision note (that hidden_states[-1] of vision models is pre-norm), but its coverage of the twelve breaking entries is not established here.1

Second, the notes are not an exhaustive changelog. They support statements of the form "the notes state X", never "X is the only change" — and several section summaries inside the release notes end mid-sentence, so their remaining detail lives in the linked pull requests rather than in the note text.

Third, nothing here replaces a trial. No install, upgrade or fine-tuning run was performed while writing this article, and the dependency floors the notes mention — the kernels minimum in particular — are not stated as numbers. Check the specific pull requests for your own expert-parallelism, audio, exporter and quantization paths before treating 5.23.0 as a drop-in replacement.

Sources

  1. 1.api.github.com (api.github.com)