Cohere has introduced two compact language models, Tiny Aya and Cohere Transcribe, designed for developers participating in the Build Small Hackathon. The hackathon requires participants to keep their total model size at or below 32 billion parameters and to deploy a Gradio app on Hugging Face Spaces. Tiny Aya, a 3.35B multilingual text generation family, covers 70+ languages, while Cohere Transcribe, a 2B automatic speech recognition model, supports 14 languages. These models are suitable for local multilingual assistants, voice interfaces, and small applications for real-world use.
Tiny Aya is available in region-specific variants, including global, water, fire, and earth, each optimized for different language groups. Developers can use the GGUF format for lightweight deployment or integrate the models with frameworks like llama.cpp, Ollama, or transformers. The models also support Python workflows with llama-cpp-python for greater control. For speech recognition, Cohere Transcribe is available for local deployment, with support for 14 languages and options for customizing language settings and punctuation control. The model can handle long audio files by automatically chunking them and processing them in segments.
The guide outlines installation steps, code examples, and deployment strategies for both models, emphasizing their suitability for local use and small-scale applications. It also provides instructions for integrating the models with Gradio for user interfaces and for serving them through local endpoints. The models are Apache 2.0 licensed and are intended for developers looking to build accessible and multilingual tools without relying on large-scale cloud infrastructure.
Source: huggingface
