Tiny Aya In The Wild

Tiny Aya Expedition Drives Multilingual Innovation | Cohere

Jul 14, 2026

9 minute read

Tiny Aya in the wild: What people are building with open multilingual AI From multilingual benchmarks to on-device assistants, discover how Tiny Aya is enabling open AI research around the world.

Key takeaways Accessible multilingual models drive innovation: Tiny Aya’s open-weight, lightweight design enables builders worldwide to create practical AI solutions, from educational tools to on-device assistants, without requiring heavy hardware. Multilingual safety requires nuanced evaluation: Projects revealed that safety behavior shifts dramatically under code-mixed prompts and cross-lingual misalignment, highlighting the need for new evaluation frameworks beyond single-language benchmarks. Local deployment unlocks real-world impact: Teams developed compression techniques, synthetic data pipelines, and privacy-preserving tools to make multilingual AI usable on phones and offline, expanding access to underrepresented languages. Language understanding goes beyond translation: Research demonstrated deep semantic cross-lingual transfer, with Tiny Aya rivaling larger models in tasks like word-sense disambiguation and even showing sensitivity to programming language syntax.

Tiny Aya is Cohere Labs' answer to a familiar problem: most AI is built for a handful of languages and needs serious hardware to run. Tiny Aya is open-weight, strong across 70+ languages, and light enough to run locally, even on a phone.

Shortly after release, we launched Expedition Tiny Aya, a mentor-supported research program to put the model in the hands of builders, researchers, students, and practitioners around the world, and see what they build with it. Alongside the growing body of community projects, the research behind Tiny Aya was recently accepted to COLM 2026, underscoring the model's impact as both a practical foundation for multilingual AI and a contribution to the research community.

What follows is an overview of what emerged from the Expedition. Projects from this initiative yielded insights across four key themes: education and learning, building safer multilingual AI, accessibility and local deployment, and language understanding and processing. In each area, a shared pattern holds: meaningful AI innovation thrives with accessible models, strong mentorship, and the freedom to explore problems that matter locally. Theme 1: Education and learning Access to quality education remains a global challenge, particularly in multilingual communities. Expedition teams tackled this by creating tools that make learning more accessible and effective.

Tiny Aya Math Edition: Solving math with transparent reasoning, Katrina Lawrence, Danylo Boiko, Jing Guo, Shubham Bari, Rabius Sany — with Cohere mentor Nikolas Gritsch Teaching Tiny Aya to show its work (breaking solutions into ideas, lemmas, sketches, and final answers) didn't guarantee correct math, but it surfaced clear tradeoffs between how a model organizes its reasoning and how well it actually performs. This team explored whether structured reasoning could improve mathematical problem solving. Alongside the project, the team released a public multilingual Math Olympiad benchmark dataset , which served as the foundation for many of their experiments. The work remains ongoing: the team is continuing model training, investigating new methods through a series of research papers, and plans to publish a comprehensive paper in the coming months.

Kids Companion: Offline multilingual education, Batuhan Aktas, Yuvraj Singh, Fatih Bugra Akdogan — with Cohere mentor Viraat Aryabumi The team built a child-safe AI voice assistant that works entirely offline . They combined Tiny Aya with speech recognition and text-to-speech systems, specifically designed for young users. Beyond the prototype, they developed a multilingual benchmark for evaluating child-focused AI systems , expanding children's conversations across dozens of languages to assess safety, educational value, and age-appropriate responses. Together, the voice companion and benchmark provide both a practical tool for young learners and an open resource that can help future researchers study multilingual AI safety and performance across diverse linguistic and cultural contexts, as described in their accompanying blog post . Theme 2: Building safer multilingual AI As models become increasingly multilingual, safety cannot be treated as a single-language problem translated outward. Instead, it becomes a question of how alignment behaves under linguistic and cultural interference.

Mix, Fine-Tune, Break: Testing multilingual safety with mixed languages , Tanav Singh Bajaj, Jasmine Zou, Elena Dobre, Naufal Adityo, Srimoyee Mukhopadhyay, and Shayam Shamsi — with Cohere mentorship from Charlie Chen This team investigated how Tiny Aya's safety behavior changes under code-mixed prompting, where conversations naturally blend multiple languages, dialects, and scripts. Their experiments showed that safety evaluations can produce markedly different results when languages are mixed rather than used in isolation, revealing blind spots that may be missed by traditional multilingual benchmarks . Their work highlights how evaluating multilingual safety requires fundamentally different approaches than traditional methods, exposing important gaps in current alignment measurement frameworks.

Cross-Lingual Emergent Misalignment: Stopping safety issues crossing languages , Joana da Matta, Anu Adesina, Akanksha Devkar, Mayank Bhaskar, Ayesha Imran, and Almustapha Yusuf — with Cohere mentorship from Daniel D’souza This team explored what happens when unsafe behavior learned in one language appears in another. Through mechanistic interpretability, they found evidence that certain directions associated with misalignment may generalize between languages, though transfer strength varied significantly across language pairs. This suggests multilingual models can share underlying safety-relevant representations, but in more nuanced ways than previously thought. The team has submitted to a major conference and their work will be made public soon.

Context Robustness: Understanding cultural conflicts in multilingual AI, Ankita Maity, Van Ngo, Tanay Nagar, Henok Ademtew, and Sajag Swami — with Cohere mentorship from Nikita Moghe This team examined how multilingual models respond when cultural signals, language, and...