Cohere open-sources 2B ASR
Cohere released an open-source 2B-parameter ASR audio model that reportedly outperforms Whisper Large v3 by ~27% on Hugging Face leaderboards — a notable step for multimodal search and audio-first RAG. An open, stronger ASR model changes the calculus for enterprise search stacks that need native audio indexing. (x.com) (x.com)
Cohere released "Transcribe," a Conformer encoder–decoder ASR model announced March 26, 2026 with a 2‑billion‑parameter footprint aimed at production deployment. (cohere.com) The model currently sits atop Hugging Face’s Open ASR Leaderboard with a reported average WER of 5.42% across the benchmark suite. (venturebeat.com) Transcribe’s Hugging Face model card and community uploads show open weights and a multilingual scope covering 14 languages including English, Mandarin, Arabic, French, Spanish and Japanese. (huggingface.co) Cohere published the model under an Apache‑2.0 license and the project is already integrated into the Transformers docs and ecosystem with a release revision tagged in March 2026. (github.com) Cohere is distributing the weights on Hugging Face while also offering Transcribe through its API (temporarily free with rate limits), and community contributions have produced ONNX/quantized builds targeted at consumer‑grade GPUs. (ai-primer.com) Launch coverage highlights Cohere positioning Transcribe as a production‑oriented, self‑hostable alternative (packaged for local inference and OpenAI‑compatible serving), signaling immediate availability for teams that need native audio indexing and transcription pipelines. (getaibook.com)
