Tiny Aya Global is an open-weights 3.35 billion parameter multilingual small language model (SLM) developed by Cohere Labs. Released in February 2026 as part of the Tiny Aya family, it is optimized for strong and balanced performance across more than 70 languages, with a specific focus on supporting underserved and low-resource linguistic communities. The model is engineered to operate efficiently on consumer hardware, including laptops and mobile devices, without requiring a constant internet connection.
The model's architecture is a dense decoder-only Transformer featuring 36 layers and a massive 262k-token vocabulary designed to ensure equitable representation across diverse scripts and languages. It employs an interleaved attention mechanism that alternates between sliding window attention (with a 4,096 window size) and full global attention in a 3:1 ratio, utilizing Grouped Query Attention (GQA) for improved memory efficiency. The training process followed a Warmup-Stable-Decay (WSD) schedule on 6 trillion tokens, followed by a specialized post-training pipeline using Fusion-of-N (FUSION) synthetic data and SimMerge to align the model with human preferences while preserving global safety standards.
In terms of capabilities, Tiny Aya Global excels in translation, summarization, and cross-lingual conversational tasks. It has demonstrated notable performance on the GlobalMGSM benchmark, where it often outperforms comparable 4B-class models in low-resource African and West Asian languages. While the model maintains high language integrity—rarely defaulting to English during non-English interactions—it is primarily specialized for linguistic tasks rather than complex mathematical reasoning or code generation.