Open-weight model infrastructure

Transformers 5.23 adds Inkling 2 and LFM3 support with a unified model-export workflow

Hugging Face expands native open-weight architecture support and introduces an ExportedModel interface for moving compatible checkpoints into supported inference runtimes.

Open Model Weights published16 Nov 2026
Primary sourceHugging Face — Transformers v5.23.0 official release
Source published2026-11-13
FOLLOW OUR COVERAGE

Follow Open Model Weights on Google See more of our open-weight AI reporting in Google Search.

Add as a Preferred Source on Google
Original cinematic technology illustration showing a glowing modular transformer network with distinct export paths to three floating compute chips in navy, teal and cyan.
Original conceptual artwork of open-model architecture integration and export pathways, not a software screenshot or an inference benchmark. AI-generated editorial illustration; not a documentary photograph.

Hugging Face released Transformers 5.23.0 on November 13, expanding its supported open-weight architectures while redesigning parts of the model-export workflow. The release adds implementations for Thinking Machines Lab's Inkling 2, Liquid AI's LFM3 and LFM3-VL, Dots 3 Note Preview, IQuestCoder and InternLM3, alongside an experimental JEV decision head. These are integration changes in the Transformers software library, not newly trained checkpoints released by Hugging Face. Readers evaluating the actual models should still inspect each publisher's weights, commit revision, license and runtime requirements separately. Primary source ↗

Inkling 2 Small receives support for a multimodal mixture-of-experts architecture with 594 billion total parameters and roughly 17 billion active parameters, as described by the release maintainers. LFM3 combines short gated convolutions with grouped-query attention, while LFM3-VL extends that family to visual inputs. Transformers also documents the experimental nature of the JEV decision interface: API behavior and configuration can still change, which matters for production applications relying on stable decision scoring.

The infrastructure headline is the new ExportedModel path. Hugging Face brings backend-specific model export under a common interface, including workflows for ONNX Runtime, OpenVINO and ExecuTorch where supported. This can reduce integration friction for teams that need to take a source-verified checkpoint out of its original Python inference stack. It does not imply that every checkpoint is supported by every exporter or that any published latency gain transfers across different accelerators, sequence lengths and model architectures.

Developers should also read the breaking changes before upgrading deployed systems. Expert-parallel interfaces and selected optional audio dependencies have changed, and the library's compatibility surface remains broader than any single benchmark. For Open Model Weights, the practical significance is improved access to running and evaluating public weights: maintainers still need exact revision pinning, independently checked licensing and workload-specific performance tests to establish a reproducible deployment.

OMW REGISTRY WATCH

What this changes for the evidence layer

This is a Transformers library and export-tooling release, not a newly trained foundation-model checkpoint. Native architecture support is not proof that every model variant is independently tested or deploys identically across backends; keep source-pinned model repositories and their separate licenses.

PRIMARY SOURCE

Hugging Face — Transformers v5.23.0 official release

This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.

Open source article ↗