Best AI for translation

The two components this architecture separates

Cohere carries a model named Command A Translate, and its technical documentation explicitly labels it as its own dedicated machine translation model. The same documentation lists all 23 supported 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 same 23-entry list. For any project deploying a bilingual or multilingual site, this is the most precise piece of documentation on language coverage anywhere in this reference.

On the other side of the architecture, Meta ships Llama 4 with open weights, and its licence permits commercial use — a component that can be deployed on internal infrastructure and whose output can be sold. Its technical specification carries a closed twelve-language list: Arabic, English, French, German, Hindi, Indonesian, Italian, Portuguese, Spanish, Tagalog, Thai and Vietnamese. No Persian. No Turkish. Meta does not leave the list open with an "including," so this is a closed, definite specification.

The ranking axis for this category is therefore not output quality. It is whether your target language is registered in the model specification, and whether the access path is open at all.

The context-window inversion in these two architectures

Command A Translate has an operating window of 8,000 tokens — the shortest capacity in this entire catalogue, shorter than any other component recorded here. Llama 4 Scout, on the other end, carries a 10-million-token window, the longest recorded value.

So the component that recognizes Persian has the processing capacity of a few pages, while the component that does not recognize Persian accepts an entire book in a single call. This is not an architectural coincidence: a dedicated translation model is engineered for the sentence and paragraph unit, not for the full document.

The operational consequence of this inversion is undocumented elsewhere: with a dedicated translation model, a site translation pipeline has to run page by page, not in a single unified pass. If your operational requirement is passing a long document through in one call, you are in fact selecting a component that was not engineered for that workload, and that choice needs to be made deliberately at the project-architecture stage.

Why licence openness carries 30 points in this category

In the other categories of this reference, licence openness is a secondary component in the scoring architecture. Not in this one.

The reason is direct: for a user based in Iran, the exact layer of the infrastructure that is unusable in practice is the hosted API. Cohere own commercial agreement explicitly names Iran in its Restricted Location definition. So a component whose weights are published under a licence that permits commercial use is a genuinely operational part of this architecture; the rest are documentation about a tool that cannot be deployed.

This same criterion exposes a distinction that a binary open/not-open flag hides. Aya Expanse publishes open weights but operates under a CC BY-NC licence — meaning that using it for a paid client-site translation job puts you in breach. Llama 4 permits commercial use but under Meta own proprietary licence, with a 700-million monthly-active-user ceiling and a mandatory "Built with Llama" notice requirement. Mistral Large 3 ships under Apache 2.0, the freest licence in this entire table, but does not register Persian in its specification.

Four criteria deliberately excluded from this rubric

A translation-quality score. Cohere describes Command A Translate as its "state-of-the-art machine translation model" without publishing any citable figure. No other vendor publishes one either. A column with no value for any candidate has no place in the rubric architecture.

Price. Of the seven candidates in this table, only two carry a published price. Cohere prints no rate for its current-generation models, Mistral publishes no model-level rate at all, and Meta ships weights with no attached price. A column with only two filled cells pulls the other five candidates down in rank for no technical reason.

Translation direction. Translation quality is not symmetric: the Persian-to-English path and the English-to-Persian path are not operationally equivalent, and every professional translator knows this. No vendor in this reference publishes separate data per direction, so this page makes no claim about it either. A language list is a set, not a directional pair.

And the context window, a criterion that was provisioned into the rubric on a trial basis and then removed — we document the reason here because it applies to future categories too. The ranking engine normalizes every numeric column between the candidates minimum and maximum. Here the minimum window is 8,000 tokens and the maximum is 10 million, a gap of three orders of magnitude. The practical result was that 256,000 tokens scored 0.4 out of a possible 15, and 128,000 tokens scored 0.2, while the real-world gap between 8,000 and 256,000 is a 32-fold difference. A column that renders a 32-fold difference as "0.4 versus 0.2" is not measuring that difference; it produces a systematic error. That is why the figure was removed from the table while the analysis remained in the prose above.

Two unranked rows in this architecture

Two candidates sit below the table because the ranking engine could not compute a rank for them, both for the same reason: the vendor has not published a language list for that specific model.

One is Command A+, Cohere own flagship product line. Cohere documentation records a "world-class translation capability" claim for this model but publishes no language list and no definite figure for it. There is a performance claim, with no data behind it to cite.

The other is Mistral Large 3, which carries the freest licence in this entire table: Apache 2.0, open weights, full commercial-use permission — exactly the property this category weights second highest. Yet Mistral publishes neither a language list nor a country list for Large 3, and the partial list it does provide is explicitly stated as non-exhaustive. The model that was the best licence option in this table was excluded from ranking purely for lack of these two documents.

This is not hidden, because that transparency is what makes the table trustworthy: a candidate with no documented evidence gets no rank, even where our working hypothesis is that its performance is good.

What this table does not measure

Every cell in the Persian column on this page states only that the vendor registered that language in its specification. No actual output has been read or evaluated.

This is the same methodological limit the voice-generation category carries, and it is the honest boundary of this entire methodology. We read and cite officially published lists; assessing how natural these models Persian output actually reads is a separate task, and we make no claim to have done it.