What is Tiny Aya Earth?
Tiny Aya Earth is a 3.35-billion-parameter, instruction-tuned multilingual language model from Cohere and Cohere Labs. An instruction-tuned model is trained to respond to natural-language directions, so it can be used for tasks such as answering questions, translating text, summarizing documents, and generating conversations without requiring a separate task-specific model.
The model belongs to Cohere's Tiny Aya family. The family includes a globally balanced model and regional variants intended to improve coverage for particular language communities. Tiny Aya Earth is the regional variant focused on languages across Africa and West Asia, while still supporting a total of 70 languages.
Its relatively small size is an important part of its positioning. Compared with much larger general-purpose language models, a 3.35B-parameter model can be more practical to run on local or edge hardware, although the available research does not establish a universal hardware requirement or performance guarantee. Tiny Aya Earth is therefore aimed at applications where multilingual coverage, deployment control, and efficiency matter more than frontier-level reasoning.
Languages and core capabilities
Tiny Aya Earth accepts text and produces text. Cohere documents it for multilingual understanding, translation, target-language generation, summarization, conversation, and cross-lingual tasks. Its language coverage includes Arabic, Persian, Urdu, Turkish, Hebrew, Amharic, Hausa, Igbo, Malagasy, Shona, Swahili, Wolof, Xhosa, Yoruba, and Zulu, along with languages from Europe, South Asia, East Asia, and Southeast Asia.
The regional focus is useful when an application must work across languages that are less consistently supported by mainstream models. For example, a system could use Tiny Aya Earth to translate customer messages between English and Swahili, summarize material written in Arabic, generate multilingual educational content, or provide a conversational interface spanning several African and West Asian languages. These examples describe suitable application patterns rather than guaranteed quality levels for any individual language.
Tiny Aya Earth uses an autoregressive transformer architecture with sliding-window and global-attention layers. In practical terms, it generates text sequentially while combining nearby context with selected broader context. Cohere's model information specifies an 8K-token context length and an 8K maximum output length.
Technical specifications at a glance
| Specification | Details |
|---|---|
| Provider | Cohere and Cohere Labs |
| Model family | Tiny Aya |
| Parameters | 3.35 billion |
| Model type | Instruction-tuned multilingual language model |
| Supported languages | 70 |
| Context length | 8,192 tokens |
| Maximum output | 8,192 tokens |
| Input modality | Text |
| Output modality | Text |
| API identifier | tiny-aya-earth |
| Open-weight checkpoint | CohereLabs/tiny-aya-earth |
| Open-weight license | CC-BY-NC with Cohere Labs acceptable-use requirements |
The 8K context limit is the amount of text the model can consider in a request, including the prompt and other conversation or document content. The 8K maximum output is a separate ceiling for generated tokens. Actual usable length can depend on the API or inference software being used.
API access and local deployment
Cohere lists tiny-aya-earth as a live model available through its Chat API. This is the simplest route for developers who want hosted inference without operating model-serving infrastructure. The supplied research does not provide a model-specific input price or output price, so a precise per-token cost cannot be stated here. Users evaluating hosted deployment should check Cohere's current pricing and account terms rather than assume that a general API rate applies to this model.
Cohere Labs also publishes the model weights through its Hugging Face organization. The open-weight release can be used with Transformers, vLLM, SGLang, compatible local applications, and quantized GGUF variants. These options allow experimentation on privately controlled infrastructure and may suit offline or edge scenarios where sending text to a hosted API is undesirable.
The open-weight release is published under a CC-BY-NC license with additional Cohere Labs acceptable-use requirements. That makes it suitable for research, evaluation, and non-commercial prototyping, but it should not automatically be treated as permitted for commercial production. Organizations planning commercial use should review the exact license and usage conditions and contact Cohere if clarification is needed.
Modalities, tools, reasoning, and coding
Tiny Aya Earth is text-only. It does not have documented native image, audio, or video input or output. It should therefore be evaluated as a multilingual text model, not as a multimodal assistant that can inspect images, transcribe audio, or generate media.
The supplied model information does not document native web search, function calling, tool execution, structured output, JSON mode, batch processing, or prompt caching for Tiny Aya Earth. An application could potentially build surrounding software that performs these functions, but that would be an application-layer feature rather than a verified native capability of the model.
It is also not documented as a reasoning specialist or coding specialist. The available editorial evaluation assigns relatively modest comparative scores for reasoning and coding, but these are editorial estimates rather than provider-published benchmark results. In practical terms, Tiny Aya Earth is better matched to translation, multilingual generation, summarization, and conversation than to difficult multi-step analysis, advanced software engineering, or dependable autonomous research.
Main strengths and trade-offs
The clearest strength is its combination of regional language focus and compact size. Many applications need to serve users in multiple languages but do not need a very large model for every request. Tiny Aya Earth provides a focused option for those workloads, particularly where African and West Asian languages are central to the product.
- Regional multilingual coverage: The model supports 70 languages and specifically targets African and West Asian language use cases.
- Compact deployment profile: Its 3.35B-parameter size is more compatible with local, edge, and research deployment than very large models, subject to available hardware and serving configuration.
- Two access paths: Developers can use Cohere's hosted Chat API or examine and deploy the open-weight checkpoint.
- Useful core tasks: Translation, summarization, conversation, target-language generation, and cross-lingual processing are the model's documented focus.
- Research flexibility: Open weights enable local experimentation and downstream adaptation without requiring every test to run through a hosted endpoint.
The trade-off is breadth and depth. Tiny Aya Earth is not presented as a frontier general-purpose model, and the research does not establish that it matches larger systems on difficult reasoning, coding, or factual research. Its language list also should not be interpreted as equal quality across all 70 languages. Lower-resource language performance can vary by language, domain, prompt, and evaluation method.
A smaller model may also produce fluent but incorrect statements. Translation and summaries should be checked when mistakes could affect legal, medical, financial, educational, or public-service outcomes. The model's multilingual fluency is useful, but fluency alone is not evidence that an answer is factually correct.
Best use cases
Tiny Aya Earth is a sensible candidate when the primary requirement is multilingual text processing with a preference for regional coverage and manageable deployment requirements. Suitable uses include:
- Translation between English and African or West Asian languages
- Multilingual customer-support conversations
- Summarizing reports, messages, or educational material
- Cross-lingual search or content transformation when paired with application-level retrieval systems
- Local or offline language assistants where hosted processing is unsuitable
- Research into lower-resource languages and multilingual model adaptation
- Generating text for prototypes that need a smaller model footprint
For production systems, developers should test the exact languages, writing styles, and subject areas they intend to support. A model that performs well in one language or domain may not perform equally well in another.
When to choose Tiny Aya Earth
Choose Tiny Aya Earth when African or West Asian language coverage is more important than advanced reasoning, native tool use, or multimodal input. It is particularly attractive when a team wants to compare hosted inference with local deployment, or when a compact open-weight model is preferable for research and non-commercial prototyping.
A larger general-purpose model may be more appropriate for complex planning, advanced coding, long chains of reasoning, broad factual research, or workflows that require built-in tool orchestration. A dedicated speech or vision model is a better choice for audio transcription or image understanding. An embedding, reranking, or search-specific model may also be more suitable when the main task is retrieval rather than generation.
Within the Tiny Aya family, the regional focus is the key reason to select Earth. A globally balanced sibling may be a better comparison point when the application has no particular African or West Asian language priority, but the supplied research does not provide a detailed head-to-head benchmark between those variants.
Limitations and licensing considerations
Tiny Aya Earth has several limitations that should be considered before deployment. It is text-only, has an 8K context and 8K output ceiling, and lacks documented native web browsing, tool use, structured output, batch processing, and caching. It is not positioned as a frontier reasoning or coding model. The model's regional specialization is valuable, but performance should be validated separately for each important language and domain.
The open-weight license is another practical constraint. CC-BY-NC generally points toward non-commercial use, and Cohere Labs adds acceptable-use requirements. Hosted API use and local-weight use may have different contractual conditions, so organizations should review the relevant terms for their chosen deployment route.
Bottom line
Tiny Aya Earth is a focused multilingual model for organizations and researchers that need text generation across 70 languages, with particular interest in African and West Asian language applications. Its 3.35B-parameter design, 8K context, hosted API availability, and open-weight release give it a useful balance of accessibility and deployment flexibility. Its strongest case is efficient multilingual work, not frontier reasoning, advanced coding, multimodal interaction, or autonomous tool-based workflows.