TinyAya: Multilingual AI for the Other 7 Billion

AI for the Rest of the World

The artificial intelligence revolution has a language problem. The frontier models that dominate headlines — GPT-4, Claude, Gemini — are designed primarily for English speakers. They perform well in a handful of other high-resource languages: French, German, Spanish, Chinese, Japanese. But for the majority of the world’s 7,000+ languages — and crucially, for many of the languages spoken by billions of people in Africa, the Middle East, South Asia, and Southeast Asia — these models range from mediocre to unusable.

On February 17, 2026, Cohere Labs launched Tiny Aya at the India AI Summit — a family of open-weight multilingual models built around a 3.35-billion-parameter base that covers more than 70 languages and runs offline on consumer hardware, from laptops to smartphones. The family includes five models: Tiny Aya Base (the pretrained foundation), Tiny Aya Global (instruction-tuned across 67 languages with balanced performance), and three regional specialists — Tiny Aya Earth for African and West Asian languages, Tiny Aya Fire for South Asian languages, and Tiny Aya Water for Asia Pacific and European languages.

Tiny Aya is small by current standards. At 3.35 billion parameters, it is roughly two orders of magnitude smaller than frontier models. It cannot match Claude or GPT-5 on complex reasoning tasks in English. But that comparison misses the point entirely. For a farmer in Kenya seeking agricultural advice in Swahili, a student in Bangladesh studying in Bangla, or a small business owner in Algeria navigating government services in Arabic, Tiny Aya offers something that no frontier model does: useful AI assistance in their own language, on hardware they can actually access, without needing an internet connection.

The Architecture of Accessibility

Tiny Aya’s design reflects a series of deliberate trade-offs between capability and accessibility. The model uses an auto-regressive transformer architecture with a hybrid attention mechanism — sliding window attention for the first three layers (window size 4,096) and global attention on the fourth layer — optimized with Rotary Position Embeddings (RoPE) and an 8K context window for both input and output.

The tokenizer is a critical component and one of Tiny Aya’s most significant technical achievements. Most frontier models use tokenizers designed primarily for English, which fragment non-English text into many small tokens. A single Arabic or Hindi word might require three to five tokens in a model with an English-centric tokenizer, compared to one or two in a multilingual-optimized tokenizer. This fragmentation increases computational cost, reduces the effective context window, and degrades generation quality. Tiny Aya’s tokenizer was specifically redesigned to reduce fragmentation across scripts, achieving what Cohere describes as the most efficient tokenization across the vast majority of evaluated languages.

The training data curation was perhaps the most labor-intensive component. The Aya initiative — a global open-science effort organized through Cohere’s research lab, Cohere Labs — mobilized over 3,000 researchers, linguists, engineers, and native speakers from 119 countries to assemble training data that included not just web text but curated examples of high-quality writing, instruction-following demonstrations, and safety-aligned responses across dozens of languages. For some low-resource African and Southeast Asian languages, this required creating training data nearly from scratch.

Remarkably, the entire model was trained on just 64 NVIDIA H100 GPUs — a modest setup by current standards that underscores the efficiency-first philosophy driving the project. Post-training included supervised fine-tuning and preference training for safety and helpfulness alignment.

The result is a model that, while not matching frontier models in any single language, provides useful functionality across a breadth of languages that no frontier model can match. Across the 70+ supported languages, Tiny Aya handles conversational interaction, question answering, text summarization, translation, and simple reasoning tasks at a level that is practically useful for everyday applications.

The Frontier Model Gap

Tiny Aya’s release throws into sharp relief a growing tension in the AI industry: the widening gap between the capabilities available to English speakers and those available to everyone else.

Frontier models are trained primarily on English data because English dominates the internet. Research from Stanford has documented how large language models systematically exclude non-English speakers, creating what researchers call a “digital language divide.” Languages spoken by hundreds of millions of people — Swahili, Yoruba, Tamil, Punjabi, Amharic — account for fractions of a percent of web content. Hausa, spoken by 80 million people, receives only 0.0036% representation in common training datasets. Models trained on web-scale data inevitably reflect these proportions, performing well in English and progressively worse in less-represented languages.

The problem is compounding. As AI tools become more deeply integrated into education, healthcare, commerce, and government services, the quality of AI in a given language increasingly determines the quality of digital services available to speakers of that language. A student with access to a fluent AI tutor in English has a qualitatively different learning experience than a student whose AI tutor barely understands their language. A doctor using AI-assisted diagnostic tools in French receives better support than a doctor using the same tools in Hausa.

This dynamic threatens to create a new form of inequality where access to AI-powered services depends not on economic resources or technical infrastructure but on the language one speaks. Research from Johns Hopkins and Brookings has shown that multilingual AI systems often reinforce rather than reduce these biases, amplifying the dominance of English while sidelining minority languages.

Tiny Aya is not a complete solution to this divide. A 3.35B model cannot match the depth of one two orders of magnitude larger. But it represents a different philosophy of AI development — one that prioritizes breadth of access over depth of capability, and that treats linguistic inclusion as a first-order design goal rather than an afterthought.

Running AI on Consumer Hardware

One of Tiny Aya’s most significant practical achievements is its ability to run on consumer-grade hardware. At 3.35 billion parameters with appropriate quantization, the model runs offline on modern laptops, and independent testing by the Futurum Group recorded the model achieving 32 tokens per second on an iPhone 17 Pro — fast enough for real-time conversational use.

This has profound implications for deployment in contexts where cloud computing is expensive, unreliable, or unavailable. Large swaths of the developing world have limited or intermittent internet connectivity. Cloud-based AI services that require constant server communication are impractical in rural areas of sub-Saharan Africa, South Asia, and Southeast Asia where cellular coverage is spotty and bandwidth is expensive. A model that runs locally on a device eliminates the dependency on cloud connectivity entirely.

Local deployment also addresses privacy concerns that are particularly acute in certain cultural and political contexts. Users in authoritarian countries may be unwilling to send sensitive queries to foreign cloud servers. Medical professionals in regions with strict data protection requirements may not be able to use cloud-based AI tools for patient-related queries. A locally running model keeps data on the device, eliminating the privacy risks associated with cloud AI.

The economic implications are significant. Cloud AI inference is expensive — each query to a frontier model costs the provider fractions of a cent, which adds up to billions of dollars annually at scale. These costs are ultimately borne by users through subscription fees or advertising. A locally running model has no per-query cost after initial deployment, making it dramatically more affordable for cost-sensitive markets.

The models are available for download on HuggingFace, Kaggle, and Ollama, with deployment supported through the Cohere Platform — lowering the barrier for developers worldwide to build applications on top of Tiny Aya.

Benchmark Performance: Small but Capable

Despite its compact size, Tiny Aya demonstrates competitive performance against models of similar and even larger scale. On the WMT24++ translation benchmark, Tiny Aya Global outperforms GEMMA3-4B — a model with slightly more parameters — in 46 out of 61 evaluated languages. The gap is particularly pronounced in low-resource languages where Tiny Aya’s specialized training data gives it a clear advantage.

On mathematical reasoning, the differences are even more striking. In the GlobalMGSM benchmark for African languages, Tiny Aya achieved 39.2% accuracy, significantly outperforming GEMMA3-4B at 17.6% and QWEN3-4B at just 6.25%. This suggests that Tiny Aya’s multilingual training data curation — particularly the contributions from native speakers in underrepresented language communities — produces real capability gains that raw model scale alone cannot replicate.

The model does have clear limitations. Chain-of-thought reasoning tasks remain challenging, factual knowledge gaps are more pronounced in low-resource languages, and quality varies across the 70+ supported languages — high-resource languages like French and Arabic perform more consistently than languages with less training data. These are honest trade-offs in a model designed for breadth rather than depth.

The regional variants help address uneven performance. Tiny Aya Fire, optimized for South Asian languages including Bengali, Hindi, Punjabi, Urdu, Gujarati, Tamil, Telugu, and Marathi, delivers stronger performance in those languages than the Global variant. Similarly, Tiny Aya Earth focuses resources on African languages including Swahili, Hausa, Yoruba, Igbo, and Amharic, producing more culturally attuned responses with better named-entity handling for those regions.

The Community Behind the Model

Tiny Aya would not exist without the Aya community — a global network of over 3,000 researchers, linguists, engineers, social scientists, and native speakers from 119 countries who contributed to the model’s training data, evaluation, and quality assurance. Organized through Cohere Labs, this community represents one of the most ambitious collaborative efforts in AI history.

The contribution model was deliberately inclusive. Native speakers of underrepresented languages — many of whom are not AI researchers — were recruited to create instruction-following demonstrations in their languages. These demonstrations were essential for instruction tuning the model. The community also provided quality evaluations, testing the model in their native languages and providing feedback on accuracy, fluency, and cultural appropriateness.

This community model highlights a potential path for AI development that does not depend entirely on the resources of large corporations. By mobilizing native speakers as contributors rather than treating them as passive consumers, the Aya project created value that could not be replicated by any amount of compute — the nuanced linguistic and cultural knowledge that only native speakers possess.

The Aya initiative has been building toward this release for years, with earlier milestones including the Aya 101 model covering 101 languages. Tiny Aya represents the distillation of that accumulated knowledge into a form factor optimized for real-world deployment rather than benchmark maximization.

What Comes Next for Multilingual AI

Tiny Aya is a milestone, but the journey toward genuinely equitable multilingual AI is long. The 70+ languages covered by the model represent a fraction of the world’s linguistic diversity. Many of the world’s most endangered languages — spoken by small communities with limited digital presence — remain entirely absent from AI systems.

The technical challenges of expanding coverage are significant. For many of the world’s 7,000+ languages, insufficient digital text exists to train even a small model. Some languages have no standardized written form, making text-based AI approaches fundamentally inadequate. Audio-based AI — speech recognition and synthesis — may be more appropriate for oral-tradition languages, but these technologies face their own data scarcity challenges.

The licensing model also raises questions. Tiny Aya is released under CC-BY-NC — non-commercial use only — with Cohere’s Acceptable Use Policy. While this makes the model freely available for research and community projects, commercial applications require a Cohere Platform agreement. This is a pragmatic business decision, but it means that the organizations most likely to deploy AI at scale in developing markets — mobile carriers, banks, government agencies — will need commercial licenses.

These questions do not have easy answers. But Tiny Aya demonstrates that the answer is not simply to wait for frontier models to get better at non-English languages — a process that has been agonizingly slow. The answer may instead lie in purpose-built multilingual models, community-driven data curation, and a philosophy of AI development that measures success not by benchmark scores in English but by real-world utility across the full spectrum of human language.

Sources & Further Reading

• Cohere Labs Launches Tiny Aya, Making Multilingual AI Accessible — Cohere Blog
• Cohere Launches a Family of Open Multilingual Models — TechCrunch
• Tiny Aya Global Model Card — HuggingFace (CohereLabs)
• The Aya Initiative: Building AI for Underrepresented Languages — Cohere Labs
• How AI Is Leaving Non-English Speakers Behind — Stanford Report
• How Language Gaps Constrain Generative AI Development — Brookings