Cohere Labs launches Tiny Aya, a multilingual 3.35B model for local

Cohere Labs, the research arm of enterprise‑AI startup Cohere, unveiled Tiny Aya—a 3.35 billion‑parameter multilingual model designed to run locally on devices such as smartphones while maintaining strong performance across more than 70 languages [1].

Key takeaways

• Tiny Aya is an open‑weight model with 3.35 B parameters, targeting efficient on‑device inference [1].
• The model claims competitive results on verification, translation, and mathematical inference tasks, especially for low‑resource languages [1].
• Four specialized variants—Fire, Earth, Water, and Global—focus on different regional language groups [1].
• Benchmark comparisons show Tiny Aya outperforming comparable lightweight models (e.g., Google Gemma 3‑4B) on West Asian and African language tasks [1].
• The model is immediately available on Hugging Face Spaces and Cohere’s platform for public testing [1].

Multilingual performance in a compact package

Cohere Labs built Tiny Aya with the explicit goal of “maintaining reliability across many languages while remaining efficient enough for local use,” according to the company’s announcement [1]. The model supports over 70 global languages and delivers strong results on standard multilingual benchmarks, rivaling larger open‑weight models in tasks such as verification, translation, and mathematical inference. In head‑to‑head tests against four lightweight competitors—including Google’s Gemma 3‑4B and Mistral AI’s 3‑3B—Tiny Aya achieved the highest scores for West Asian and African language pairs [1].

To address regional linguistic nuances, Cohere released four specialized versions—Tiny Aya Fire, Earth, Water, and Global—each tuned for specific language clusters across five world regions [1]. Post‑training efforts focused on 67 languages, and tokenization efficiency analyses show Tiny Aya achieving the most efficient tokenization for most languages tested [1]. The model’s small footprint enables it to run on edge devices, a rarity for multilingual models of this capability.

Open access and next steps

Tiny Aya is distributed as an open‑weight model, with code and weights hosted on Hugging Face and directly accessible via Cohere’s platform [1]. Researchers and developers can experiment with the model through a dedicated Hugging Face Space, facilitating rapid adoption and community feedback. Cohere’s broader multilingual initiative includes earlier releases such as the Aya 101 model, which covers over 100 languages, and the newer Aya Expanse family targeting higher‑parameter models for research use [2][3][4]. Tiny Aya therefore represents the latest step in Cohere’s strategy to democratize multilingual AI by offering increasingly capable models that can be deployed locally.

Why it matters

Tiny Aya demonstrates that high‑quality multilingual AI can be delivered in a compact form factor, lowering barriers for developers who need on‑device language capabilities—particularly in regions with limited connectivity or compute resources. By outperforming other lightweight models on low‑resource language tasks, Tiny Aya helps address the persistent gap in AI services for many non‑English speakers. The open‑weight release invites broader community involvement, potentially accelerating innovation in multilingual applications ranging from translation tools to localized assistants. As Cohere continues to expand its Aya family with larger models, Tiny Aya sets a benchmark for balancing performance, efficiency, and accessibility in the evolving landscape of multilingual .