**Added:**  
- New release **v5.12.0** (marked as Latest) – published 12 Jun.  
  - **New model additions:**  
    - **MiniMax-M3-VL:** Vision-language model pairing a CLIP-style vision tower with a mixed dense/sparse MoE text backbone (3D rotary embeddings, SwiGLU-OAI gated experts, block-sparse attention).  
    - **PP-OCRv6:** Lightweight OCR system with MetaFormer-style backbone and structural reparameterization, offered in medium/small/tiny tiers.  
    - **Parakeet-RNNT:** Fast Conformer encoder + RNN-T decoder for speech recognition (LSTM prediction network, greedy transducer decoding).  
  - **Bugfixes and improvements:** Over 20 commits covering CI OTEL exports, stop string matching for byte-fragment tokens, DiffusionGemma training and documentation, security gate automation for PRs, tokenizer fixes, and various trainer/attention updates.  
  - **Significant community contributions:** @ArthurZucker (MiniMax-M3-VL) and @eustlb (Parakeet-RNNT) listed.  
  - **Contributors:** kashif, tarekziade, and 14 others.

**Changed:**  
- The **“Latest” badge** was removed from the v5.11.0 release entry (it now reads “Release v5.11.0” without the badge).  
- **Pagination** updated from “1 2 3 4 5 … 25 26” to “1 2 3 4 5 … 26 27” to accommodate the new release.

**Removed:**  
- The **v5.6.0 release entry** was removed entirely. That release had included:  
  - Model additions (OpenAI Privacy Filter, QianfanOCR, SAM3-LiteText, SLANet)  
  - Breaking changes (internal rotary_fn registration)  
  - Enhancements to `transformers serve`, vision, parallelization, tokenization, cache, and audio  
  - A large set of bugfixes and improvements contributed by earino, kashif, and 68 others.