Regional and specialist models

The map has more on it than six labs

The families in the previous lesson are the ones written about. They are not the whole ecosystem, and for a large share of the world they are not the right starting point either. Two other groups matter: models built for particular languages, and models built for particular domains.

The reason to care is mechanical rather than patriotic. A general model trained overwhelmingly on English carries two disadvantages in another language, and both are measurable. It has seen less of the language, so it knows less. And its tokenizer splits that language into more tokens, so every request costs more and every context window holds less. A model built for the language fixes both at once, which is why a 7B Indic-tuned model can beat a general 70B on Hindi summarisation while losing badly on English reasoning.

Language-focused families worth knowing

India. AI4Bharat, at IIT Madras, has released the most useful public work: IndicTrans2 for translation across 22 scheduled languages, IndicBERT, and speech models covering languages that commercial systems ignore. Sarvam AI builds Indic-tuned models on open bases and releases weights. Krutrim has released Indic models. The practical point for anyone working here is that translation, transliteration and speech in Indian languages are better served by these than by a general chat model, and the gap is largest in exactly the languages with the least web text.

South East Asia. SEA-LION from Singapore’s AI Singapore covers Bahasa Indonesia, Thai, Vietnamese, Tamil and others with a tokenizer designed for them.

Arabic. Jais, from G42 and MBZUAI, was the first serious open Arabic-centric family. Falcon, from TII in Abu Dhabi, is Apache 2.0 and general rather than Arabic-only.

Europe. Teuken and the OpenEuroLLM effort target the 24 official EU languages, with the explicit aim of not being an English model with translation bolted on.

Multilingual generalists. Cohere For AI’s Aya covers a very large number of languages and is good. Note the licence carefully: Aya models have shipped under CC-BY-NC, which is non-commercial. This is the single most common licence surprise in this part of the map, because the model is excellent and the announcement reads like any other open release.

Domain-specialised models

Medical. Meditron, OpenBioLLM, and several biomedical tunes exist and score well on medical question benchmarks. Treat every one of them as a research artefact. A model that answers medical exam questions well is being measured on a format, not on patient safety, and none of these releases has regulatory clearance for clinical use anywhere. The honest use is literature triage and drafting for a qualified reader.

Legal. SaulLM and similar legal tunes. The same caution applies with an extra edge: legal answers are jurisdiction-specific, and a model tuned largely on one country’s corpus will answer confidently about another.

Science and code. Protein language models, chemistry models, and the code families that get their own lesson next.

Enterprise-shaped. IBM’s Granite is Apache 2.0, deliberately unexciting, and tuned for tool use and cheap serving. Nvidia’s Nemotron models are re-tuned derivatives of other bases, optimised for Nvidia serving stacks.

How to decide whether a specialist is worth it

Run the general model and the specialist on twenty of your own examples. Three outcomes, and each has a clear response.

• The specialist wins clearly. Usually happens on translation, transliteration, speech, and anything where the language itself is the difficulty. Take the win.
• They tie. Usually happens on classification and extraction, where the task carries most of the signal and the language carries little. Take the general model, because its ecosystem is deeper and it will be maintained longer.
• The specialist wins on the target language and loses everywhere else. Common. This is an argument for routing: the specialist behind the tasks that need it, the general model behind everything else. Two models on disk is not a problem; two models nobody measured is.

Check the tokenizer as part of this. Encode a paragraph of your language in both and compare token counts. A specialist that produces 40% fewer tokens for the same text is 40% cheaper per request and holds 40% more context, before any quality argument.

The failure to avoid is picking a specialist because it is from your region rather than because it measured better. The reason these models often win is real and mechanical; the way to know it applied to your task is to run it.