v5.13.0
📦 transformersView on GitHub →
✨ 8 features🔧 8 symbols
Summary
Release v5.13.0 introduces a significant number of new model architectures across language, vision, and speech domains, including Kimi 2.x, MiMo-V2-Flash, Nemotron ASR variants, Qwen3 ASR, ZAYA1, VideoPrism, and RADIO.
✨ New Features
- Added support for Kimi K2.5, K2.6, and K2.7 multimodal agentic models.
- Added support for MiMo-V2-Flash, an MoE language model supporting up to a 256K context window.
- Added support for Nemotron 3.5 ASR, a multilingual speech recognition model with configurable streaming chunk sizes.
- Added support for Nemotron ASR Streaming, an English speech recognition model optimized for low-latency streaming.
- Added support for Qwen3 ASR, an automatic speech recognition model with a Whisper-style encoder and Qwen3 decoder, including a forced aligner.
- Added support for ZAYA1, a 760M active parameter MoE language model from Zyphra.
- Added support for VideoPrism, a general-purpose video encoder for diverse video understanding tasks.
- Added support for RADIO (Reduce All Domains Into One), a family of vision foundation models producing image-level summary embeddings and dense spatial features.