The Pulse

Transformers 5.23 adds support for Inkling 2 Small

Hugging Face’s Transformers 5.23 supports Thinking Machines Lab’s text-and-image model Inkling 2 Small. The release also changes model exports and adds support for several other model families.

AI.info Team · · fact-checked · against 3 sources ↓ · Published 13 Nov 2026

Transformers 5.23 supports Inkling 2 Small, a Thinking Machines Lab model that takes text and images and generates text. The Nov. 13 release also adds other model families and changes how exported models are saved and loaded, according to Hugging Face's release notes.

Inkling 2 Small is intended for English and other languages, as well as multiple coding languages. Its language backbone has 48 mixture-of-experts decoder layers, each routing tokens to 8 of 384 experts. The model has 594 billion total parameters and 17 billion active parameters.

Text and images, but not audio

Unlike the first-generation Inkling Small, the new model does not accept audio input, Hugging Face says in a separate technical post. The earlier model has 276 billion total parameters and 12 billion active parameters.

The new model retains features of the Inkling decoder, including learned relative attention, short causal convolutions and mixture-of-experts layers. It adds block attention residuals, gated attention with a learned sink, and dynamic key and value convolutions.

For images, the model uses a vision transformer based on the Perception Encoder and projects the resulting features into the language model. The vision encoder splits images into 20-by-20-pixel patches and merges each 2-by-2 group, producing one language-model embedding for each 40-by-40-pixel region.

The model supports a context window of up to 1 million tokens. Thinking Machines released it on Nov. 5 under an Apache 2.0 license, with open weights available for download.

Running the model

Hugging Face's example for Inkling 2 Small uses the same Auto classes as the earlier Inkling model: AutoProcessor and AutoModelForMultimodalLM. The example loads the model, prepares a message containing an image and a text question, and generates a text response.

Thinking Machines distributes the weights in two checkpoint formats. Its model card says the main NVFP4 checkpoint requires a GPU cluster with at least 500GB of aggregated VRAM; the alternative MXFP8 checkpoint requires at least 800GB of aggregated VRAM.

The model can also be accessed through Thinking Machines' Tinker fine-tuning platform and through third-party inference providers. The model card lists several inference frameworks, including Hugging Face, for running the weights directly on GPU hardware.

Other additions in 5.23

The release also adds support for LFM3 and LFM3-VL, Liquid AI model families that include a text model and a vision-language model. LFM3-VL takes images and videos at their native resolution alongside text.

Another addition is JEV, a decision head that scores single-token labels. JEV and its decision pipeline are experimental, and their APIs, configuration and behavior may change; the release notes advise caution when using them in production.

The update changes exports so that every export returns an ExportedModel, which can be called like the original model, pushed to the Hub and loaded on the backend's runtime. The same export API covers Dynamo, ONNX Runtime, OpenVINO and ExecuTorch.

The release also includes vision fixes, including corrections for Qwen VL positions when a text cache is reused with a new image and for image handling in Emu3. Image and video classification pipelines now respect the problem_type setting.