Transformers 5.18.0 adds four model families, breaks six APIs

Hugging Face's release adds Nemotron 3 diarization and NemotronH Omni alongside six backwards-incompatible fixes.

Hugging Face has shipped Transformers 5.18.0. It adds four new model families and flags six breaking changes across the library used to run most open-weight models in production. The release notes list the usual long tail of bugfixes. But the headline additions are a streaming diarization model and three multimodal systems.

Nemotron 3 Diarization, an open-weight model, answers the practical question of who spoke when in real audio. It supports both streaming and offline inference, and handles up to eight speakers. It orders its outputs by each speaker's first appearance in the clip. A single checkpoint covers latency profiles from an 80 millisecond input buffer up to a 30.4 second offline-style buffer, with output frame resolution configurable in multiples of 10 milliseconds. Chunked inference removes any cap on audio duration.

NemotronH Omni, also from NVIDIA, pairs the NemotronH hybrid Mamba-Transformer language model with a RADIO vision encoder and an optional Parakeet-based sound encoder. It projects image, video and audio inputs into the language model's embedding space. A single autoregressive model then reasons across all four modalities at once.

NAVER's HyperCLOVAX Vision V2 combines the HyperClovaX language backbone with a Qwen2.5-VL vision encoder. It adds chain-of-thought reasoning through built-in <think> tokens, alongside text, image and video support. GTE, a BERT-style bidirectional encoder described in the mGTE paper, swaps absolute position embeddings for RoPE and adds a gated MLP. The same architecture already underpins Alibaba's gte-*-v1.5 and gte-multilingual-* checkpoints and Snowflake's snowflake-arctic-embed-m-v2.0. This is, in effect, a reference implementation catching up to models already in wide use.

The six breaking changes are narrower in scope but will bite specific users. They touch ROCm's gpt-oss attention routing to aiter-flash-attn, a faster DETR image processor, and a remap of indexer layer_type. They also include a fix to vLLM's video token counting for the Transformers backend, an update to the DINO model family, and a version bump to the kernels dependency. Anyone pinning Transformers in a production pipeline that touches those six areas should check the linked pull requests before upgrading.

Written from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.