CohereLabs / tiny-aya-l2-thinker

Model Card for Tiny Aya L2-Thinker

Model Summary

Cohere Labs Tiny Aya L2-Thinker is an open-weights research release of a 3.35 billion parameter multilingual reasoning model. It is trained to think in the same language as the user prompt (L2 thinking) before writing the final answer.

The bundled chat template does this automatically: every user turn is wrapped with a language-forcing prefix and a /think suffix, and generation is primed with <|START_THINKING|> . These are already included in the chat template.

Developed by: Cohere and Cohere Labs

• Point of Contact: Cohere Labs
• License: CC-BY-NC , requires also adhering to Cohere Lab's Acceptable Use Policy
• Model: tiny-aya-l2-thinker
• Model Size: 3.35B
• Context length: 32K input

For the broader Tiny Aya family, see tiny-aya-global , tiny-aya-base , and the Tiny Aya collection .

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "CohereLabs/tiny-aya-l2-thinker"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype="auto")

messages = [
    {"role": "user", "content": "Explica en español qué significa la palabra japonesa 'ikigai'."},
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
    return_dict=True,
).to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=32768,
    do_sample=True,
    temperature=0.6,
    top_p=0.95,
)

print(tokenizer.decode(outputs[0], skip_special_tokens=False))

apply_chat_template(..., add_generation_prompt=True) produces a prompt of this shape (the user message is wrapped for you):

<BOS_TOKEN><|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|># System Preamble
...Aya system + default preamble...
<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>Think in the same language as the prompt. [USER MESSAGE] /think<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|><|START_THINKING|>

The model then writes a thinking trace between <|START_THINKING|> and <|END_THINKING|> , followed by the user-facing answer between <|START_RESPONSE|> and <|END_RESPONSE|> .

You can also pass prior thinking back into the conversation:

messages = [
    {"role": "user", "content": "How many r's are there in strawberry?"},
    {
        "role": "assistant",
        "thinking": "Count the letters: S-T-R-A-W-B-E-R-R-Y. Three r's.",
        "content": "There are 3 r's in strawberry.",
    },
    {"role": "user", "content": "Now do the same for blueberry."},
]

Chat template behavior

The tokenizer chat template always:

1. Inserts the Tiny Aya system prompt:

# System Preamble
You are in contextual safety mode. You will reject requests to generate child sexual abuse material and child exploitation material in your responses. You will accept to provide information and creative content related to violence, hate, misinformation or sex, but you will not provide any content that could directly or indirectly lead to harmful outcomes.

Your information cutoff date is June 2024.

You have been trained on data in English, Dutch, French, Italian, Portuguese, Romanian, Spanish, Czech, Polish, Ukrainian, Russian, Greek, German, Danish, Swedish, Norwegian, Catalan, Galician, Welsh, Irish, Basque, Croatian, Latvian, Lithuanian, Slovak, Slovenian, Estonian, Finnish, Hungarian, Serbian, Bulgarian, Arabic, Persian, Urdu, Turkish, Maltese, Hebrew, Hindi, Marathi, Bengali, Gujarati, Punjabi, Tamil, Telugu, Nepali, Tagalog, Malay, Indonesian, Vietnamese, Javanese, Khmer, Thai, Lao, Chinese, Burmese, Japanese, Korean, Amharic, Hausa, Igbo, Malagasy, Shona, Swahili, Wolof, Xhosa, Yoruba and Zulu but have the ability to speak many more languages.

# Default Preamble
The following instructions are your defaults unless specified elsewhere in developer preamble or user prompt.
- Your name is Aya.
- You are a large language model built by Cohere.
- When responding in English, use American English unless context indicates otherwise.
- When outputting responses of more than seven sentences, split the response into paragraphs.
- Prefer the active voice.
- Use gender-neutral pronouns for unspecified persons.
- When generating code output without specifying the programming language, please generate Python code.

1. Prefixes every user turn with Think in the same language as the prompt. simulating training data.
2. Appends /think to every user turn.
3. Starts the assistant generation with <|CHATBOT_TOKEN|><|START_THINKING|> .

Model Details

Input: Text only.

Output: Model generates text, including an explicit thinking trace.

Model Architecture: Auto-regressive transformer in the Tiny Aya / Cohere family. After pretraining, this checkpoint is supervised-fine-tuned for multilingual reasoning so that the thinking language follows the prompt language.

Languages covered: Amharic, Arabic, Basque, Bengali, Bulgarian, Catalan, Chinese, Czech, English, Filipino, Finnish, French, German, Greek, Hausa, Hebrew, Hindi, Hungarian, Igbo, Indonesian, Irish, Italian, Japanese, Javanese, Khmer, Korean, Lithuanian, Malay, Maltese, Norwegian, Persian, Polish, Punjabi, Russian, Slovak, Swahili, Tamil, Telugu, Thai, Turkish, Ukrainian, Urdu, Vietnamese, Yoruba, and Zulu.

Context Length: Tiny Aya L2-Thinker supports a context length of 32K.

Usage and Limitations

Intended Usage

Tiny Aya L2-Thinker is meant for multilingual reasoning and conversational use, especially when the thinking trace should stay in the user's language rather than defaulting to English. Intended applications include multilingual math and reasoning, open-ended generation, and research on target-language reasoning.

Limitations

As with any language model, outputs may contain incorrect or outdated statements. Thinking traces can be long; cap max_new_tokens appropriately. Lowest-resource languages may show more variability than high-resource ones.

Model Card Contact

For errors or additional questions about details in this model card, contact [email protected] .

Terms of Use

This model is governed by a CC-BY-NC License (Non-Commercial) and also requires adhering to Cohere Lab's Acceptable Use Policy . If you are interested in commercial use, please contact Cohere’s Sales team.