Cohere AI Labs has released Tiny Aya, a family of small language models (SLMs) that redefines multilingual performance. While many models scale by increasing parameters, Tiny Aya uses a 3.35B-parameter architecture to deliver state-of-the-art translation and generation across 70 languages.
The release includes 5 models: Tiny Aya Base (pretrained), Tiny Aya Global (balanced instruction-tuned), and three region-specific variants—Earth (Africa/West Asia), Fire (South Asia), and Water (Asia-Pacific/Europe).
The Architecture
Tiny Aya is built on a dense decoder-only Transformer architecture. Key specs include:
- Parameters: 3.35B total (2.8B non-embedding)
- Layers: 36
- Vocabulary: 262k tokenizer designed for equitable language representation.
- Attention: Interleaved sliding window and full attention (3:1 ratio) with Grouped Query Attention (GQA).
- Context: 8192 tokens for input and output.
The model was pretrained on 6T tokens using a Warmup-Stable-Decay (WSD) schedule. To maintain stability, the team used SwiGLU activations and removed all biases from dense layers.
Advanced Post-training: FUSION and SimMerge
To bridge the gap in low-resource languages, Cohere used a synthetic data pipeline.
- Fusion-of-N (FUSION): Prompts are sent to a ‘team of teachers’ (COMMAND A, GEMMA3-27B-IT, DEEPSEEK-V3). A judge LLM, the Fusor, extracts and aggregates the strongest components of their responses.
- Region Specialization: Models were finetuned on 5 regional clusters (e.g., South Asia, Africa).
- SimMerge: To prevent ‘catastrophic forgetting’ of global safety, regional checkpoints were merged with the global model using SimMerge, which selects the best merge operators based on similarity signals.
Performance Benchmarks
Tiny Aya Global consistently beats larger or same-scale competitors in multilingual tasks:
- Translation: It outperforms GEMMA3-4B in 46 of 61 languages on WMT24++.
- Reasoning: In the GlobalMGSM (math) benchmark for African languages, Tiny Aya achieved 39.2% accuracy, dwarfing GEMMA3-4B (17.6%) and QWEN3-4B (6.25%).
- Safety: It holds the highest mean safe response rate (91.1%) on MultiJail.
- Language Integrity: The model achieves 94% language accuracy, meaning it rarely switches to English when asked to reply in another language.
On-Device Deployment
Tiny Aya is optimized for edge computing. Using 4-bit quantization (Q4_K_M), the model fits in a 2.14 GB memory footprint.
- iPhone 13: 10 tokens/s.
- iPhone 17 Pro: 32 tokens/s.
This quantization scheme results in a minimal 1.4-point drop in generation quality, making it a viable solution for offline, private, and localized AI applications.
Key Takeaways
- Efficient Multilingual Power: Tiny Aya is a 3.35B-parameter model family that delivers state-of-the-art translation and high-quality generation across 70 languages. It proves that massive scale is not required for strong multilingual performance if models are designed with intentional data curation.
- Innovative Training Pipeline: The models were developed using a novel strategy involving Fusion-of-N (FUSION), where a ‘team of teachers’ (like Command A and DeepSeek-V3) generated synthetic data. A judge model then aggregated the strongest components to ensure high-quality training signals even for low-resource languages.
- Regional Specialization via Merging: Cohere released specialized variants—Tiny Aya Earth, Fire, and Water—which are tuned for specific regions like Africa, South Asia, and the Asia-Pacific. These were created by merging regional fine-tuned models with a global model using SimMerge to preserve safety while boosting local language performance.
- Superior Benchmark Performance: Tiny Aya Global outperforms competitors like Gemma3-4B in translation quality for 46 of 61 languages on WMT24++. It also significantly reduces disparities in mathematical reasoning for African languages, achieving 39.2% accuracy compared to Gemma3-4B’s 17.6%.
- Optimized for On-Device Deployment: The model is highly portable and runs efficiently on edge devices; it achieves ~10 tokens/s on an iPhone 13 and 32 tokens/s on an iPhone 17 Pro using Q4_K_M quantization. This 4-bit quantization format maintains high quality with only a minimal 1.4-point degradation.
Check out the Technical details, Paper, Model Weights and Playground. Also, feel free to follow us on Twitter and don’t forget to join our 100k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
