Summary: Cohere Labs released Tiny Aya on February 17, 2026 as an open-weight multilingual family covering more than 70 languages in 3.35 billion parameters. The article emphasizes low-resource languages, laptop-scale local deployment, quantizations, and a hosted inference rate of roughly 133 tokens per second. It is licensed CC-BY-NC-4.0, and the model card describes weaker chain-of-thought, math, and low-resource factual performance.
On February 17, 2026, Cohere Labs — the research arm of enterprise AI company Cohere — released Tiny Aya, an open-weight family of multilingual models that squeezes coverage of more than 70 languages into 3.35 billion parameters, a size small enough to run on a laptop. TechCrunch covered the launch, which landed on the sidelines of the India AI Impact Summit in New Delhi.
Why it matters: multilingual capability has largely lived inside frontier models gated behind APIs, and quality drops fast once you step outside high-resource languages. Tiny Aya bets the other way — open weights, a deliberate emphasis on low-resource languages, and local deployment as the design goal, at a size consumer hardware can genuinely run. For teams building outside the English-first internet, that combination simply was not available before at this price point: free.
3.35B Parameters, 70+ Languages
The base model weighs in at 3.35 billion parameters, with BF16 weights totaling roughly 6.7 GB. Coverage stretches across European, Indic, East and Southeast Asian, African, and Middle Eastern language families — 67 languages tagged on the Hugging Face model card — with a deliberate skew toward low-resource ones: Zulu, Wolof, Igbo, Lao, Javanese, and Burmese all appear. That skew is the whole point. High-resource languages are already well served by every frontier API; the languages on this list are exactly where those APIs quietly fall back to translation pipelines and second-rate generation quality.
Architecturally, the model combines sliding-window attention (a 4096 window with RoPE position encoding) with a single global-attention layer, and offers 8K of context on both the input and output side. This hybrid is the standard trade-off when a modest parameter budget has to serve long multilingual documents: the sliding window keeps compute predictable, while the lone global layer retains enough long-range signal for cross-sentence coherence in languages the model encountered less of during training.
TinyAya-Global: The Instruction-Tuned Release
TechCrunch singled out TinyAya-Global, a variant fine-tuned to follow user instructions for applications requiring broad language support. Per the model card, it is tiny-aya-base aligned through supervised fine-tuning plus preference training, and it is positioned as the family’s best-balanced option across languages and regions.
Genuinely Laptop-Scale: Deployment
The model card states the design goal plainly: local deployment under realistic compute constraints. The practical routes run through transformers, vLLM, SGLang, and Docker Model Runner, and the community has already shipped eight quantizations. If self-hosting is not your plan, Cohere’s hosted inference serves roughly 133 tokens per second — handy for prototyping before you commit to running it yourself. Set against frontier models that weigh hundreds of gigabytes, this size turns “offline multilingual assistant” from a demo into a shippable product.
The deployment story matters more than the raw parameter count. A 6.7 GB BF16 model fits on ordinary laptops and small cloud instances, and the published quantizations push the footprint lower still. That is the gap between a model you merely evaluate and a model you actually put in front of users in markets where bandwidth limits and privacy expectations both argue for on-device inference.
Licensing and the Capability Ceiling
The license is CC-BY-NC-4.0 (non-commercial) plus Cohere’s acceptable use policy; the repository sits behind a contact form, and commercial use goes through Cohere sales. The capability limits are stated without spin: the card lists a single eval — GPQA Diamond at 28.28%, far below frontier reasoning models — and openly notes weaker chain-of-thought and math performance (MGSM-class tasks) plus the risk of factual errors, especially in low-resource languages. The positioning is clear: a workhorse for generation, translation, and cross-lingual summarization, not a reasoning model. If your product needs multi-step planning, keep a frontier model in the loop and use Tiny Aya where the job is language coverage at zero marginal cost.
The Aya Lineage, and Two Lessons for Builders
Tiny Aya grows out of Cohere’s Aya program, an open-science multilingual initiative whose earlier research model covered 101 languages. Staging the debut at the India AI Impact Summit was a deliberate choice of stage: the markets with the most to gain from cheap, multilingual, local AI are exactly the Global South users the summit brought together.
Two lessons for product teams. First, unit economics in non-English markets shift: a 3.35B local multilingual base means zero per-token bills and data that never leaves the device — cost and privacy improve at the same time. Second, the non-commercial license is the startup red line: prototype freely, but settle the license before you ship — the most common last mile for open-weights products.