Hugging Face Transformers has released version 5.16.1, a noteworthy update that introduces support for GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series. Alongside this new model, the release includes a pair of small patches aimed at restoring backward compatibility and addressing a security concern.
GLM-5.3-Flash: A Leaner, More Capable Multimodal Model
GLM-5.3-Flash breaks new ground for the GLM family by being the first model to natively handle multiple modalities. Despite its massive 320 billion total parameters, only 18 billion are active during inference, which allows it to outperform its predecessor, GLM-5.2, across various benchmarks and real-world tasks at one-tenth of the price. On coding and agentic benchmarks, it approaches the performance of Claude Opus 4.8.
The model was built from a newly trained base, with an architecture and training recipe redesigned for capability and efficiency. For the first time in the GLM series, it uses a hybrid architecture that combines sparse and linear attention, sharply reducing serving costs for long contexts while preserving precise long-context capabilities. It also incorporates Manifold-Constrained Hyper-Connections (mHC) to further improve scaling efficiency. These innovations, together with a 30 trillion-token multimodal pre-training corpus, enable GLM-5.3-Flash to deliver more intelligence with less compute.
Small Patch Fixes
In addition to the new model, version 5.16.1 includes two small but important patches. The first restores backward compatibility for the tensor-parallel API, ensuring that users who relied on previous behavior won’t experience disruptions. The second pins a Hugging Face kernel commit and corrects repository paths for ESMFold2, addressing a security concern. These fixes are part of the ongoing effort to keep the library stable and secure for self-hosted deployments.
