What is Tiny Aya?
Tiny Aya comes from Cohere Labs, Cohere's research arm. Tiny Aya Global is the all-round chat model; Earth, Fire and Water lean towards African and West Asian, South Asian, and European and Asia-Pacific languages; and a base model is there for fine-tuning. All read and write text only, and their knowledge stops at June 2024.
It is current: the four chat models are live on the Cohere API, and Cohere later added two 32K reasoning models. Next to Gemma or Qwen models of the same size, it trades reasoning and a longer context for steadier quality across less common languages.
What are Tiny Aya L2-Thinker and En-Thinker?
They are later additions to the Tiny Aya family: two 3.35B reasoning models that think step by step before answering, with a 32K-token window.
- L2-Thinker
- Writes its reasoning in the same language as your prompt, trained for 44 languages plus English.
- En-Thinker
- Reasons in English and then answers in your language, for the same 44 languages plus English.
- Base 32K
- The pretrained base with a 32K window that both Thinkers are built on, for fine-tuning and long-context research.
- Where to get them
- Open weights on Hugging Face under the same non-commercial licence; Cohere's API docs list only the four original chat models.
Who is Tiny Aya best for?
Tiny Aya is best for researchers, non-profits and developers building offline or phone-based tools in lower-resourced languages, where steady quality across 70 languages matters more than a commercial licence.
Offline translation
It runs on laptops and phones with no cloud connection; Cohere measured about 10 tokens a second on an iPhone 13 with the 4-bit file.
Lower-resourced languages
Quality holds steadier across languages with little web text, and on African-language math it scored 39.2% against 17.6% for Gemma3-4B in Cohere's tests.
Multilingual NLP research
Open weights, a base model, a 32K base and a detailed technical report make it a starting point for fine-tuning new languages.
Regional non-commercial projects
The Earth, Fire and Water variants focus on one region each, and Cohere reports up to 5.5 ChrF points better translation than Global for South Asian languages.
What's included with Tiny Aya?
Tiny Aya comes as five 3.35B models (a base, Global and three regional variants) with official GGUF files, Cohere API access for the chat models, and two newer reasoning models.
- Five models, one size
- A pretrained base, the balanced Global chat model and three regional chat models, all at 3.35B parameters.
- Regional variants
- Earth for Africa and West Asia, Fire for South Asia, and Water for Europe and Asia-Pacific.
- Official GGUF files
- 4-bit, 8-bit and 16-bit builds of each chat model run in llama.cpp, Ollama, LM Studio and Jan.
- Cohere API
- The four chat models are on Cohere's Chat endpoint, so you can call them with the Cohere SDK.
- Thinker models
- L2-Thinker reasons in the language of your prompt and En-Thinker reasons in English, both with a 32K-token window.
- Efficient tokenizer
- It splits text in most scripts into fewer tokens than other small models, which cuts memory and compute on local hardware.
How much does Tiny Aya cost?
Tiny Aya is free: the weights download at no cost under a non-commercial CC-BY-NC licence, and the Cohere API runs the chat models free on a rate-limited trial key, with no published production price.
- All five 3.35B models on Hugging Face and Kaggle
- Official GGUF files from 2.03 GB
- Commercial use needs a deal with Cohere's sales team
- Models tiny-aya-global, -earth, -fire and -water on the Chat endpoint
- Trial keys are rate limited and capped at 1,000 API calls a month
- No production price published for Tiny Aya
Cohere's pricing page lists no Tiny Aya rate; production keys are billed pay as you go, and Cohere sends commercial users of Tiny Aya to its sales team. A free demo runs on Hugging Face Spaces, so you can try the models before downloading them. Full pricing on cohere.com. We checked these prices on .
How do readers rate Tiny Aya?
No one has rated Tiny Aya yet. Be the first.
Sign in to rate. One rating per account.
Tiny Aya pros and cons
Tiny Aya handles less common languages well for its size and runs on a phone, but its non-commercial licence, 8K context and weaker reasoning limit where you can use it.
Pros
- Handles less common languages well for its size: in Cohere's tests Tiny Aya Global out-translated Gemma3-4B in 46 of 55 WMT24++ languages and led same-size models on open-ended writing.
- Small enough for a phone: the official 4-bit file is 2.14 GB and loses 1.4 points of quality against the full model, by Cohere's measure.
- Reddit users see it as a ready-made translator across many language pairs, useful without the work of fine-tuning a general model.
- Open weights in five variants plus official GGUF builds, and the chat models also run on the Cohere API.
Cons
- Non-commercial licence: CC-BY-NC 4.0 plus Cohere's acceptable use policy rules out business use without a deal, the most common complaint on Reddit.
- The 8K-token context is small next to Gemma 4's 128K or Qwen3.5's 262K, and several users called it limiting.
- Weaker at reasoning and math: Cohere's own report puts it 2 to 10 points behind same-size rivals on GlobalMGSM, and users found Gemma and Qwen better in Finnish and Asia-Pacific languages.
- TranslateGemma 4B, a translation-only model, scores about the same on average in Cohere's report and wins in most languages; Tiny Aya leads in 27 of 66 on Flores.
What are the best Tiny Aya alternatives?
Three tools on Recaply do Tiny Aya's job. Qwen3.5 Small and Gemma 4 are open source, so you can run them on your own server.
TranslateGemmaFreeChoose it when you only need translation and want to use it in a product: Google's Gemma terms have no non-commercial clause, it comes in 4B, 12B and 27B, and it also translates text in images, though it covers 55 languages and takes 2K tokens of input.
Qwen3.5 SmallOpen SourceChoose it when you need a commercial licence, long inputs and stronger reasoning: its 0.8B to 9B models are Apache 2.0, read 262K tokens and cover 201 languages and dialects.
Gemma 4Open SourceChoose it when you want a phone-sized model you can ship: E2B and E4B are Apache 2.0, take 128K tokens and read images and audio as well as text.
Questions about Tiny Aya
What is Tiny Aya?
Tiny Aya is Cohere Labs' family of 3.35B-parameter open models trained on 70 languages, built to translate and write in lower-resourced languages on laptops and phones, for research and non-commercial use.
Who is Tiny Aya for?
Tiny Aya is best for researchers, non-profits and developers building offline or phone-based tools in lower-resourced languages, where steady quality across 70 languages matters more than a commercial licence.
Is Tiny Aya free?
Yes. Tiny Aya has a free Open weights plan: all five 3.35B models on Hugging Face and Kaggle.
Is there an open source alternative to Tiny Aya?
Yes. Qwen3.5 Small and Gemma 4 are open source tools that do the same job, so you can run your own.
What are the best Tiny Aya alternatives?
The closest alternatives on Recaply are TranslateGemma, Qwen3.5 Small and Gemma 4.
Where can I try Tiny Aya?
On its official website, docs.cohere.com.
Can I use Tiny Aya commercially?
Not without a deal. Tiny Aya's weights are under CC-BY-NC 4.0 plus Cohere Labs' acceptable use policy, and Cohere's trial API keys may not be used for production or commercial work. Cohere asks anyone who wants commercial use to contact its sales team.
Does Tiny Aya run on a phone?
Yes. Cohere ran the 4-bit Q4_K_M build, a 2.14 GB file, through MLX on iPhones: about 10 tokens a second on an iPhone 13 and 32 on an iPhone 17 Pro, with an average quality drop of 1.4 points against the full model.
Which Tiny Aya model should I use?
Start with Tiny Aya Global for balanced quality across all languages. Switch to Earth for African and West Asian languages, Fire for South Asian languages or Water for European and Asia-Pacific ones, and use the base model only if you plan to fine-tune it.
Is Tiny Aya good at translation?
For its size, yes, by Cohere's tests: Tiny Aya Global beat Gemma3-4B in 46 of 55 languages on WMT24++, and the regional variants add up to 5.5 ChrF points for South Asian languages. The translation-only TranslateGemma 4B scores about the same on average.
Can I run Tiny Aya in Ollama?
Yes. Cohere's official GGUF files load in Ollama straight from Hugging Face with ollama run hf.co/CohereLabs/tiny-aya-global-GGUF:Q4_K_M, and the same files work in llama.cpp, LM Studio and Jan.
Our verdict
Tiny Aya is a good pick for offline, non-commercial work in less common languages, and a weak one for anything else.
Cohere's tests show steadier translation and writing across 70 languages than same-size Gemma and Qwen models, and it runs on a phone, but it trails them on reasoning and stops at 8K tokens.
- Skip it for
- Commercial products, long documents and math-heavy work; Gemma 4 or Qwen3.5 Small cover those under Apache 2.0.
- Check first
- Test your own language pairs against TranslateGemma before you commit, since results vary a lot by language.
Gemini 3.5 Live Translate
OpenL
Excel Translator
Violin
ChatGPT Translate
SYSTRAN