The African-language models you are allowed to ship

If you are picking a model for an African language, quality is the second question. The first is whether you are permitted to put it in a product, and for this particular set of languages the answer has been “no” more often than anywhere else in the field.

This is a practical guide to which side of that line each option sits on, for text and for speech, with the reasoning you would need to check us.

The trap: the best models were the blocked ones

Two models have sat in our catalogue for months marked as unservable rather than quietly deleted, because the pattern they represent is the point.

Aya Expanse 8B is among the strongest open models for African languages. It is CC-BY-NC-4.0. Non-commercial. You may research with it and you may not sell anything built on it.

Meta MMS covers more African languages for speech than anything else that existed at the time. Also CC-BY-NC-4.0. Same answer.

That is the shape of the problem. The models trained specifically to serve languages the big labs skipped were released by research groups, and research groups default to non-commercial. So the better a model was at Yoruba or Amharic, the more likely you were forbidden from building a business on it.

Two licences, and only one of them is about the weights

This is the distinction that costs people the most time, so it is worth being precise.

The weights licence governs the model. Apache 2.0 or MIT grants use, modification, distribution, sublicensing and sale, with no revenue cap and no clause about offering the model as a service.

A supplier’s terms of service govern the API you bought. Those are a different instrument and they are usually narrower. DeepInfra bars competitive use without prior written consent. Kimi’s OpenPlatform terms bar renting, lending, selling or sub-licensing without authorisation. DeepSeek uses near-identical wording.

Going direct to the model’s creator does not escape that second clause. Buying an API is buying an API, whoever sells it. What escapes it is the weights licence: if the weights are Apache 2.0, you may run them yourself and nobody’s permission is involved.

Both of those checks are real, and they are separate. Passing one does not pass the other.

What changed: Omnilingual ASR

In November 2025 Meta released Omnilingual ASR: 1,600 languages, 4.3 million hours of training audio, character error rate under 10% on 78% of them. Model sizes from 300M to 7B.

Under Apache 2.0. Not CC-BY-NC. It is the successor to MMS, and Meta changed the licence.

That single decision retires the blocker described above for speech. The published per-language figures include 2.8% on Swahili, 3.5% Hausa, 5.7% Amharic, 9.1% Igbo, 4.9% Xhosa. For comparison, we measured Whisper on the same languages last week and got 20%, 43%, 98%, 70% and 92%.

It also covers all eleven African languages that Whisper and Qwen3-ASR have no language code for at all, which is the failure no amount of accuracy tuning addresses.

The catch is real and worth stating plainly: no serverless provider hosts it. Using it means running it yourself. The 300M CTC variant is small enough that this is a much smaller undertaking than self-hosting a chat model, and ONNX conversions already exist, but it is still a GPU you are responsible for. We do not serve it today.

Simba-S from UBC-NLP is the other one worth knowing: 43 African languages, built on SeamlessM4T-v2, released CC-BY-4.0. Commercial use permitted with attribution. Same catch, nobody hosts it.

Where the text side stands

Better than the speech side, and it has been for a while. Of the 46 models we serve today, 36 are Apache 2.0 or MIT. Llama and Gemma carry community licences with acceptable-use riders and revenue thresholds you are extremely unlikely to hit, but read them rather than assuming.

The gap on the text side is not licensing, it is cost. A token is not a word, and the same sentence costs more in Swahili than in English on every tokenizer we have measured, by between 1.47x and 7.69x depending on the model. We wrote that up separately in how to choose a model for an African language, and the per-model multipliers are on /models.

How to check a model yourself

Four steps, in this order, because each one can end the process.

1. Find the weights licence on the model card, not the product page. Apache 2.0 and MIT are unconditional. CC-BY-NC ends it. CC-BY-SA is share-alike, which may attach to a model you fine-tune on it. Anything else, read it.
2. Read your supplier’s terms for a resale or competitive-use clause. This is separate from step 1 and is the one people skip.
3. If the model was fine-tuned, check what it was fine-tuned on. A permissive base does not launder a non-commercial dataset. Several strong African-language fine-tunes on Hugging Face are built on corpora with their own restrictions.
4. Check the language is actually supported rather than only listed. For speech this is concrete: send the language code and see whether the API accepts it. Eleven African languages return a 422 on every speech model we can buy.

What we do with this

Every model in our catalogue carries its licence and a commercialUse flag, and the flag is a gate rather than a label: a model that fails it cannot be returned by the resolver, so it cannot be served by accident. Aya and MMS are still in that file, marked blocked, with the reason written next to them. GET /v1/models never returns them.

The two datasets that matter for what comes next are both new and both usable. Google’s WAXAL covers 19 languages for speech recognition and 17 for synthesis, mostly CC-BY-SA-4.0. The Gates-funded African Next Voices dataset is 9,000 hours across 18 languages under CC-BY-4.0, with restrictions on synthesis and voice cloning. Read the share-alike on the first one before training anything you intend to keep closed.

None of this makes the models good yet. It makes them legal, which for these languages has been the harder of the two problems.