Best AI for translation

The model that knows your language and the model you can run are not the same model

Cohere has a model called Command A Translate and its documentation calls it, in those words, Cohere machine translation model. In the same place it names its 23 languages one at a time: English, French, Spanish, Italian, German, Portuguese, Japanese, Korean, Chinese, Arabic, Russian, Polish, Turkish, Vietnamese, Dutch, Czech, Indonesian, Ukrainian, Romanian, Greek, Hindi, Hebrew and Persian.

All four languages this site publishes in sit inside that one list of 23. For anyone running a four-language site, that is the most precise sentence about translation in this reference.

Now the other side. Meta ships Llama 4 with open weights and a licence that permits commercial use, so it is a model you can take, run on your own hardware and sell the output of. Its language list has twelve entries: Arabic, English, French, German, Hindi, Indonesian, Italian, Portuguese, Spanish, Tagalog, Thai and Vietnamese. No Persian. No Turkish. And Meta does not leave the list open with an including, so it is a closed list.

The axis of this category is therefore not quality. It is whether the model names your language, and whether you can reach it at all.

The context window inversion every site translation runs into

Command A Translate has an 8,000 token window. That is the shortest in this entire catalogue, shorter than any other model recorded here. Llama 4 Scout has ten million tokens, the longest.

So the model that knows Persian holds a few pages, and the model that does not name Persian reads a book in one go. That is not a coincidence: a dedicated translation model is built for the sentence and the paragraph, not for the document.

The practical consequence, which nobody seems to have written down: with a dedicated translation model you translate a site page by page, not in one pass. If you want a whole long document to go in one call, you are choosing a model that was not built for translation, and that choice should be deliberate.

Why licence openness carries 30 here

In the other categories of this reference, open weights are a nice extra. Not here.

The reason is plain: for a reader in Iran, the link in the chain that does not work is the hosted API. Cohere own commercial agreement writes Iran by name into its Restricted Location definition. So a model whose weights are published, under a licence that permits commercial use, is genuinely a tool here, and the rest are articles about tools.

And this criterion exposes a difference that a simple has-open-weights boolean hides. Aya Expanse publishes weights but under CC BY-NC, so translating a client site with it and charging for the work puts you in breach. Llama 4 permits commercial use but under Meta own licence, with a 700 million monthly user ceiling and a mandatory Built with Llama notice. Mistral Large 3 is Apache 2.0, the freest licence in this table, and it does not name Persian.

Four criteria deliberately left out

A translation quality score. Cohere writes that Command A Translate is its state of the art machine translation model and publishes no number. Neither does anyone else here. A column with a figure for no candidate is not a column.

Price. Of the seven candidates in this table, two have one. Cohere prints no rate for any current-generation model, Mistral prints no per-model rate at all, and Meta hands over weights with no price attached. A column with two cells filled drags five candidates down for no reason.

Direction. Translation quality is not symmetric: Persian into English is not the same job as English into Persian, and every translator knows it. No maker in this reference publishes anything per direction, so this page says nothing about it. A language list is a set, not a pair.

And the context window, which we tried as a column and took out. The reason is worth writing down because it is a lesson in itself. The ranking engine normalises every numeric column between the lowest and the highest candidate. Here the lowest window is 8,000 tokens and the highest is ten million, three orders of magnitude apart. The result was that 256,000 tokens scored 0.4 out of 15 and 128,000 scored 0.2, while in practice the gap between 8,000 and 256,000 is thirty-two fold and is exactly what shapes a real project. A column that renders a thirty-two fold difference as 0.4 against 0.2 is not measuring it. So the number came out of the table and the argument stayed in the prose above.

The two rows that could not be ranked at all

Two candidates sit below the table because the engine could not rank them, and both for the same reason: their maker publishes no language list for that particular model.

One is Command A+, Cohere own flagship. Cohere documentation says this model has world-class translation capabilities and then gives no language list and no definite number for it. So there is a translation claim and nothing to cite.

The other is Mistral Large 3, which carries the freest licence in this whole table: Apache 2.0, open weights, full commercial use. Exactly what this category weights second highest. But Mistral publishes neither a language list nor a country statement for Large 3, and the list it does give is explicitly not exhaustive. The model that was the best option in this table on licence stayed out of the table for want of those two lists.

We do not hide this, because it is what makes the table worth trusting: a candidate with no evidence does not get ranked, even when we suspect it is good.

What this table does not measure, said plainly

Every language cell on this page says the maker listed that language. Nobody has read the output and judged it.

That is the same limit the voice category carries and it is the honest boundary of this whole method. We read published lists and cite them; judging how natural these models Persian actually reads is a different job and we do not claim it.