v5.16.1
📦 transformersView on GitHub →
✨ 4 features🐛 1 fixes🔧 3 symbols
Summary
This release introduces GLM-5.3-Flash, a new natively multimodal model, and includes a few small fixes, including restoring backward compatibility for the tensor-parallel API.
Migration Steps
- Restore BC for the tensor-parallel API.
✨ New Features
- Introduced GLM-5.3-Flash, a natively multimodal model with 320B total parameters and 18B active parameters, offering improved performance and efficiency over GLM-5.2.
- GLM-5.3-Flash features a hybrid architecture combining sparse and linear attention for reduced long-context serving costs.
- GLM-5.3-Flash incorporates Manifold-Constrained Hyper-Connections (mHC) for improved scaling efficiency.
- GLM-5.3-Flash is trained on a new 30T-token multimodal pre-training corpus.
🐛 Bug Fixes
- Fixed kernel commit and repository paths for ESMFold2.