February 2026 monthly summary for huggingface/transformers: Delivered Qwen3.5 multimodal model support with new configurations, model architectures, and text+vision processing capabilities, plus rotary position embedding improvements for multimodal performance. Implemented auto image processor, simplified vision config, and ensured robust text config checks. Aligned changes with mainline stability, and added a test suite and documentation to support adoption and reliability.

September 2025 focused on stabilizing Qwen3-Next normalization by aligning its l2norm with the FLA library. The fix replaces the previous RMS normalization with a custom l2norm to ensure compatibility and correctness across model operations, particularly within the GDN. The change enhances reliability for downstream deployments and reduces numerical drift when integrating with external libraries. The work is tracked under commit 98a8078127b4a41d1c2d78487abf3bf3831e914b (#40842).

April 2025 monthly summary for ggml-org/llama.cpp: Delivered Qwen3 and Qwen3MoE model architectures support in the llama framework, including the registration of new model classes and updates to architecture constants to enable causal language modeling with these models. This work broadens model compatibility and prepares the ground for broader experimentation and adoption within llama.cpp.