Cohere AI Launches Tiny Aya: 3B-Parameter Multilingual Model Runs Offline on Phones

Cohere AI Launches Tiny Aya: 3B-Parameter Multilingual Model Runs Offline on Phones

Cohere AI Labs has introduced Tiny Aya, a groundbreaking family of small language models (SLMs) designed to deliver advanced multilingual capabilities without requiring cloud connectivity or high-end hardware. With a compact 3.35 billion parameter architecture, Tiny Aya supports 70+ languages for tasks including translation, text generation, and instruction following—making it the most linguistically diverse open SLM to date. Unlike conventional large language models that demand significant computational resources, Tiny Aya is optimized to run efficiently on smartphones, laptops, and other edge devices, enabling offline AI access in low-bandwidth and resource-constrained environments worldwide.

According to eWEEK, the release includes five distinct models: Tiny Aya Base (pretrained), Tiny Aya Global (balanced instruction-tuned), Tiny Aya English, Tiny Aya Spanish, and Tiny Aya Arabic. Each variant is fine-tuned for specific use cases, from general-purpose multilingual understanding to region-specific linguistic nuance. The models are fully open-source under the Apache 2.0 license, inviting developers, researchers, and organizations to deploy, modify, and extend their functionality without licensing barriers.

TechCrunch reports that Cohere’s strategic pivot toward small models reflects a growing industry shift away from the unsustainable arms race of scaling parameters. "Tiny Aya proves that performance doesn’t require bloat," said a Cohere spokesperson in an exclusive interview. "We’re not competing with trillion-parameter models—we’re redefining what’s possible with a fraction of the resources. This is AI for the global south, for rural communities, for mobile-first users who can’t rely on constant cloud access."

The implications for global equity in AI access are profound. In regions where internet infrastructure is unreliable or expensive—such as parts of Sub-Saharan Africa, Southeast Asia, and Latin America—Tiny Aya enables real-time translation of medical information, educational content, and government services directly on a user’s device. For humanitarian organizations, this means field workers can communicate with local populations in native dialects without needing to transmit sensitive data over insecure networks.

Technical benchmarks presented by Cohere show Tiny Aya Global outperforms other SLMs of similar size on the MMLU (Multi-language Multi-task Language Understanding) benchmark, achieving scores comparable to much larger models in translation and comprehension tasks. The model was trained on a curated dataset of over 1.5 trillion tokens, with a strong emphasis on underrepresented languages including Swahili, Tamil, Urdu, and Quechua—languages often excluded from mainstream AI training corpora.

While Cohere Health, a separate entity focused on healthcare AI automation, shares a similar name, it has no connection to Cohere AI Labs or the Tiny Aya project. Confusion between the two organizations has been noted by industry analysts, but Cohere AI Labs has clarified that Tiny Aya is a product of its AI research division, not its healthcare subsidiary.

Developers can now download Tiny Aya models via Hugging Face and integrate them into mobile apps using ONNX Runtime or TensorFlow Lite. Early adopters have already deployed the models for real-time voice-to-text transcription in refugee camps, multilingual chatbots for public health campaigns, and offline educational tools in remote schools.

As AI continues to centralize in corporate data centers, Tiny Aya represents a powerful counter-movement: democratized, decentralized, and linguistically inclusive. With its open architecture and minimal hardware requirements, Cohere is not just releasing a model—it’s releasing a new paradigm for global AI accessibility.