Cohere Releases TinyAya: Open-Source Multilingual Model Challenges SOTA with 3B Parameters

Cohere, a leading enterprise AI company, has launched TinyAya, a groundbreaking open-source multilingual language model with over 3 billion parameters, designed to deliver state-of-the-art (SOTA) performance across 23 languages while maintaining efficiency for deployment on modest hardware. Announced via the company’s research blog and made available on Hugging Face, TinyAya represents a significant leap in democratizing high-performance multilingual AI for developers, researchers, and organizations with limited computational resources.

According to Cohere’s technical report, TinyAya was trained on a diverse, high-quality multilingual dataset spanning over 100 languages, with focused optimization for 23 key languages including Spanish, Mandarin, Arabic, Hindi, and Swahili. The model achieves competitive results on benchmarks such as MMLU, XTREME, and XGLUE, often matching or exceeding larger models while using 70% less memory. This efficiency is achieved through architectural innovations including dynamic sparse attention and quantization-aware training, allowing TinyAya to run effectively on consumer-grade GPUs and even edge devices.

The release coincides with Cohere’s broader strategy to expand its open-model ecosystem. In parallel with TinyAya, Cohere has also released the full Aya Expanse family — its flagship multilingual models — under permissive licenses, enabling commercial use and modification. This move positions Cohere as a counterweight to proprietary AI giants, appealing to enterprises seeking transparency, data sovereignty, and customization. As noted in the company’s blog, TinyAya is not merely a research artifact but a production-ready tool, complete with fine-tuning guides, evaluation scripts, and benchmark comparisons.

Industry analysts see the launch as a strategic precursor to Cohere’s anticipated IPO. TechCrunch reported in February 2026 that Cohere generated $240 million in annual revenue, driven by enterprise contracts with Fortune 500 firms and government agencies seeking secure, private AI deployments. The company’s North and Compass platforms, which integrate its AI models into workplace productivity and business intelligence systems, have become core revenue drivers. With TinyAya, Cohere is not only enhancing its public image as an open AI advocate but also building a developer community that will, in turn, expand its enterprise adoption pipeline.

Unlike many open models that sacrifice multilingual capability for size, TinyAya demonstrates that efficiency need not come at the cost of linguistic breadth. Independent researchers have already begun replicating its results on low-resource hardware, with early adopters reporting 90% accuracy on cross-lingual question-answering tasks using only a single 24GB GPU. The model’s open licensing also invites contributions from global communities, particularly in underrepresented language regions — a key differentiator in a market dominated by English-centric models.

Cohere’s decision to release TinyAya under an Apache 2.0 license further signals its commitment to long-term ecosystem growth. The company’s emphasis on privacy, customization, and enterprise-grade security — as outlined on its official site — ensures that TinyAya can be deployed in regulated industries such as healthcare, finance, and legal services without compromising compliance.

As AI models grow increasingly massive and energy-intensive, TinyAya offers a compelling alternative: high performance without bloat. With its combination of multilingual excellence, open accessibility, and enterprise readiness, TinyAya may well redefine expectations for what a "small" language model can achieve — and could catalyze a new wave of localized, sustainable AI innovation worldwide.