Cohere Launches Tiny Aya: 3.35B Open-Weight Models Supporting 70+ Languages for Offline Use
Overview
Cohere unveiled Tiny Aya at the India AI Summit on February 17, 2026 — a family of compact, open-weight multilingual models designed to run offline on standard laptops. The release targets language accessibility in regions underserved by English-centric AI tools.
Model Specifications
- Parameters: 3.35 billion
- License: Open-weight (MIT)
- Training infrastructure: Single cluster of 64 H100 GPUs
- Languages supported: 70+
Regional Variants
Cohere released region-specific fine-tuned variants:
- TinyAya-Fire: South Asian languages — Hindi, Urdu, Bengali, Punjabi, Gujarati, Tamil, Telugu, Marathi
- TinyAya-Earth: African languages
- TinyAya-Water: Asia-Pacific, Western Asian, and European languages
Availability
Models are available on HuggingFace, Kaggle, and Ollama for local deployment, as well as through the Cohere platform API. The efficient training footprint — just 64 H100 GPUs — also positions Tiny Aya as a reference point for cost-effective multilingual model development.
Source: TechCrunch
Related Articles
The draw for LocalLLaMA was not just another coding model, but Cohere asking the local-inference crowd to test pre-release weights first.
Ollama 0.17, released February 22, introduces a new native inference engine replacing llama.cpp server mode, delivering up to 40% faster prompt processing and 18% faster token generation on NVIDIA GPUs, plus improved multi-GPU tensor parallelism and AMD RDNA 4 support.
Released April 29, 2026 under Modified MIT license, Mistral Medium 3.5 consolidates the company's chat, reasoning, and coding models into one 128B dense open-weight model with 256K context, scoring 77.6% on SWE-bench Verified.