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

Cohere Labs, the research arm of enterprise AI company Cohere, announced the release of Tiny Aya, a compact multilingual language model that can run locally on devices such as smartphones [1]. The 3.35 billion‑parameter base model supports more than 70 languages and is offered as open‑weight software, immediately available on Hugging Face and Cohere platforms [1].

Key takeaways

• Tiny Aya delivers strong multilingual performance across 70+ languages while keeping the model size small enough for on‑device inference [1].
• Four specialized variants—Tiny Aya Fire, Earth, Water, and Global—target regional language groups with tailored strengths [1].
• Benchmark tests show Tiny Aya outperforms comparable lightweight models, especially on West Asian and African language tasks [1].
• The model is part of Cohere’s broader Aya initiative, which also includes larger research models like Aya Expanse and Aya 101 [2][4].

Tiny Aya’s design and performance

Cohere Labs built Tiny Aya to balance reliability across many languages with the efficiency needed for local execution. The model’s 3.35 billion parameters place it well below larger multilingual models such as Google’s Gemma 3‑27 b, yet it achieves high quality scores on a benchmark covering 66 languages [1]. Tokenization efficiency charts indicate Tiny Aya processes most languages more efficiently than its peers [1].

The family includes four region‑focused variants—Fire, Earth, Water, and Global—each optimized for specific linguistic clusters. Post‑training on 67 languages across five world regions underpins these specializations, allowing the model to excel in tasks like verification, translation, and mathematical inference for low‑resource languages [1].

Cohere’s broader multilingual AI agenda

Tiny Aya is one component of Cohere’s larger Aya program, which began as an open‑science collaboration to advance multilingual AI. Subsequent releases include Aya Vision, Aya Expanse, and Aya 101, each scaling up parameter counts and language coverage while maintaining open‑weight accessibility [2][4]. Aya Expanse, for example, introduced 8 billion‑ and 32 billion‑parameter models covering 23 languages and was released on Kaggle and Hugging Face [4].

Cohere emphasizes transparency and community involvement, licensing its models and datasets openly to enable researchers to audit, extend, and innovate responsibly [2]. This approach aims to democratize advanced AI capabilities, reducing computational overhead and infrastructure costs for diverse research groups [2].

Why it matters

Tiny Aya’s release marks a step toward making high‑quality multilingual AI accessible without reliance on cloud infrastructure, which can lower barriers for developers in regions with limited connectivity. By offering open‑weight models that run on modest hardware, Cohere supports a broader ecosystem of applications—from local translation tools to low‑resource language research. The model’s competitive performance on West Asian and African languages suggests it could help address longstanding gaps in AI support for underrepresented linguistic communities. Future developments in the Aya family are likely to continue expanding language coverage and model efficiency, reinforcing Cohere’s role in shaping inclusive, open multilingual AI.