Cohere’s Tiny Aya: The 3.35B-Parameter Model Bringing 70+ Languages to Your Pocket

Key Takeaways

• Cohere Tiny Aya packs 3.35 billion parameters and supports 70+ languages including Hindi, Tamil, Bengali, and Marathi
• Five variants Base, Global, Earth, Fire, Water each optimized for a specific regional or deployment context
• Runs entirely offline on standard laptops and phones; no internet or cloud subscription required
• Outperforms Gemma 3-4B and Qwen 3-4B on low-resource language benchmarks (GlobalMGSM)

Most AI models demand English fluency, stable broadband, and a cloud server to function. Cohere’s Tiny Aya removes all three requirements at once. Launched February 17, 2026, at the India AI Summit, this open-weight model runs 70+ languages fully offline on a standard laptop, no internet required. What the specs, benchmark results, and real deployment data actually show is more consequential than the announcement itself.

What Is Cohere Tiny Aya?

Cohere Labs Cohere’s dedicated research division released Tiny Aya on February 17, 2026, as a family of open-weight small language models. The base model contains 3.35 billion parameters and handles more than 70 languages, with particular strength in underserved regional languages across Africa, South Asia, Asia-Pacific, West Asia, and Europe. The term “open-weight” means the model’s underlying code and weights are publicly available for anyone to use and modify.

Tiny Aya was built specifically to run on consumer hardware without an internet connection. Cohere engineered the underlying software to require less computing power than most comparable models, making offline translation and local language applications viable even on everyday devices. A technical report detailing the full training methodology is forthcoming from Cohere.

Also Read Alibaba Qwen 3.5: The Open-Weight AI Model Built for Autonomous Agents

Five Tiny Aya Variants and What Each Does

Cohere released five models in the Tiny Aya family, each built for a different deployment scope. The core design principle across all variants is consistent: strong regional linguistic grounding while retaining broad multilingual coverage, making every variant a flexible starting point for further adaptation and research.

Cohere explained the rationale directly: “This approach allows each model to develop stronger linguistic grounding and cultural nuance, creating systems that feel more natural and reliable for the communities they are meant to serve. At the same time, all Tiny Aya models retain broad multilingual coverage, making them flexible starting points for further adaptation and research.“

All five variants retain full cross-lingual capability regional specialization sharpens local performance without removing universal coverage.

Architecture: What Cohere Has Confirmed

Cohere has confirmed a limited set of architecture and training details, with a full technical report still pending. The verified facts are:

• Total parameters: 3.35 billion
• Training hardware: Single cluster of 64 Nvidia H100 GPUs
• Training scale: Described as “relatively modest computing resources” by Cohere
• On-device optimization: Software built specifically to require less computing power than most comparable models

No further architectural specifications, layer counts, vocabulary sizes, attention mechanisms, or pretraining token counts have been verified from primary sources as of this publication. This article will be updated when the official technical report is released.

Tiny Aya vs. Gemma 3: Where It Wins and Where It Falls Short

Tiny Aya and Gemma 3 target different priorities Tiny Aya optimizes for multilingual depth and offline accessibility, while Gemma 3 prioritizes context length and multimodal capability.

On the GlobalMGSM benchmark which tests translation and mathematical reasoning across diverse languages Tiny Aya outperforms both Gemma 3-4B and Qwen 3-4B specifically on low-resource African and West Asian languages. Gemma 3 retains a significant lead on context length (128K vs. 8K tokens) and is the stronger option for tasks involving image-text inputs or long documents.

On-Device Performance: Real Hardware Results

Cohere designed Tiny Aya’s software from the ground up to minimize compute requirements for on-device inference. Independent testing by Futurum Group recorded Tiny Aya achieving 32 tokens per second on an iPhone 17 Pro, a benchmark that confirms practical usability on current consumer mobile hardware.

This performance figure is significant for India specifically. A developer building a Hindi or Tamil language app can run TinyAya-Fire locally on a mid-range device without GPU acceleration, cloud API costs, or dependence on consistent broadband access. Cohere highlighted India’s linguistically diverse infrastructure as a primary motivator for the model’s offline-first architecture.

Why the India AI Summit Launch Matters

Cohere chose the India AI Summit as Tiny Aya’s global launch venue, a deliberate signal about the model’s target market. TinyAya-Fire directly addresses eight of India’s most widely spoken languages: Bengali, Hindi, Punjabi, Urdu, Gujarati, Tamil, Telugu, and Marathi. No prior small language model had covered this combination with offline capability at this parameter scale.

Cohere’s business trajectory reinforces the strategic focus. The company posted $240 million in annual recurring revenue at the end of 2025, with 50% quarter-over-quarter growth, and CEO Aidan Gomez has confirmed plans for a public offering. Tiny Aya positions Cohere as the leading multilingual AI infrastructure provider for sovereign and emerging market deployments ahead of that IPO.

How to Access and Deploy Tiny Aya

Tiny Aya is available across four platforms, with no paywall for model weights:

1. Hugging Face   Download model weights and access training and evaluation datasets
2. Kaggle   Local deployment and notebook-based experimentation
3. Ollama   Single-command local setup for offline use on personal hardware
4. Cohere Platform   Managed API access for production-scale enterprise integrations

Developers can use any of the regional variants as a starting point for fine-tuning toward more specific tasks or narrower language domains.

Limitations to Know Before Deploying

• 8K context window: Unsuitable for long-document tasks legal texts, extended reports, or multi-turn conversations beyond roughly 6,000 words
• Text-only: No multimodal capability; cannot process images, audio, or video
• Reasoning ceiling: Optimized for edge inference; complex multi-step logical reasoning favors larger models such as Cohere’s own Command R+
• Language depth variance: 70+ languages does not mean uniform quality; major Indic and African languages are primary targets, smaller dialects may perform inconsistently
• Pending technical report: Full architecture, training data, and benchmark methodology details are not yet publicly available performance claims beyond GlobalMGSM should be treated as provisional

Also Read Box Selects Cursor AI: How Enterprise Coding Platform Transformed Developer Productivity

Frequently Asked Questions (FAQs)