Translation API for 33 languages, per token

The Hunyuan MT2 models are built for one job: translating between 33 languages. A dedicated translation model costs a fraction of a general chat model and answers faster, which is what you want when the work is a product catalogue, a documentation site, support tickets or subtitles. AIFans API serves all three through the OpenAI-compatible API.

Price and what a document costs

Per 1M tokens, and the one-off cost of translating a 50,000-word document (about 70K tokens in and about as many out, at 1.4 tokens per word of English prose):

• Pay as you go per token; no subscription, no minimum spend. Failed calls are never billed.
• These are our prices, already below the model maker's own list price for every item — the models page shows both side by side.
• Online top-ups carry a payment processing fee of 5.5% (minimum $0.8), added at checkout and shown before you pay; it is not part of your credit.

Which of the three

How to call it

There is no special endpoint: it is a chat model, so you put the instruction in the system message and the text in the user message.

# pip install openai
from openai import OpenAI

client = OpenAI(api_key="YOUR_API_KEY", base_url="https://api.aifans.fan/v1")
resp = client.chat.completions.create(
    model="hy-mt2-lite",
    messages=[
        {"role": "system", "content": "Translate the user's text into Spanish. Keep the Markdown. "
                                      "Use 'panel' for 'dashboard'. Output the translation only."},
        {"role": "user", "content": "## Dashboard\n\nYour usage is updated every minute."},
    ],
    max_tokens=2000,
)
print(resp.choices[0].message.content)

Things that matter in practice

• The context window is 8K tokens and the longest answer is 4K tokens, so split long documents: one paragraph, cell or subtitle block per call, and keep the input well under 4K tokens so the translation fits in the answer.
• Send the pieces in parallel (10 to 20 at a time) rather than one long request: each key allows 100 requests a minute by default, and short calls finish in well under a second.
• Put your glossary and tone in the system message, and say “output the translation only”, so you do not pay for a preamble you throw away.
• Keep the markup in the text (Markdown, HTML tags, placeholders such as {name}) and tell the model to leave it alone; translating around the markup is what breaks layouts.
• For a whole web page or PDF in the browser instead, see the Immersive Translate guide.

When a general model is the better choice

MT2 handles translation, not rewriting. If you need a text localised rather than translated — different examples, different currency, a different sales pitch — or you need a language outside the 33, use a general model such as DeepSeek V4.1 Flash: it costs more per token but it can follow that kind of instruction.