News
ResearchOpenAI pens AI safety agreement with four firmsReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierOpenAI’s New Hacking Model Found Dozens Of Holes In Apple TechFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIResearchOpenAI pens AI safety agreement with four firmsReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierOpenAI’s New Hacking Model Found Dozens Of Holes In Apple TechFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAI
EveryLLM

Regional and language models

Models built for a language or a country: Sarvam, Krutrim, Jais, Falcon, Swallow, HyperCLOVA X. 71 models on EveryLLM so far.

Regional and language models are built for a particular language, country or region. Most large models learn mainly from English web text, so they can be weaker in other languages, use more tokens for them and miss local culture and law. Governments, universities and companies from India to the Gulf, Europe, East Asia and Africa have responded with their own models, often called sovereign AI when the state funds or directs them.

Why build one

A model trained on more Hindi, Arabic or Korean text usually writes those languages more naturally and splits them into fewer tokens, which makes it faster and cheaper to run. Builders also cite control: keeping data in the country, matching local law and values, and not depending on a few foreign suppliers. Critics point to the cost of training and ask whether adapting a strong open model would serve users better.

Programmes around the world

India's IndiaAI Mission, approved in March 2024 with an outlay of ₹10,371.92 crore, funds GPU compute and chose Sarvam AI in April 2025 to build a sovereign model; BharatGen and others joined in September 2025. The UAE backs Falcon and Jais, Saudi Arabia backs ALLaM, Switzerland released Apertus in September 2025, and South Korea named five teams for a national foundation model project in August 2025.

Built from scratch or adapted

Some regional models are trained from scratch, such as Jais, Apertus, BharatGen's Param-1 (2.9 billion parameters, Hindi and English, 2025) and Sarvam's 30B and 105B models shown in February 2026. Others start from an open model such as Llama, Mistral, Gemma or Qwen and continue training on local text, which is far cheaper. Both routes are common, and many projects have tried each.

Famous regional models

Still to be written: Clova (virtual assistant).

Further reading

The weekly digest

Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.