# Free open AI models caught up

> Free "open" AI models, which anyone can download and run, are now almost as good as the best paid ones.

Source page: https://calankii.com/en/forces/free-open-ai-models-caught-up/
Last checked: 2026-11-12
Publisher: Calankii (https://calankii.com)

## The force

There are two kinds of AI model. "Closed" models are run by a company and rented by the use. "Open" models can be downloaded, changed and run on your own computers. Open models used to be far behind. They are not any more.

The Stanford AI Index 2026 measured the gap using the public Arena rankings. It shrank from 174 points (15.2%) in May 2023 to 7 points (0.5%) in August 2024. In March 2026 it was 49 points, which the report gives as 3.4% in one place and 3.3% in another ([Stanford AI Index 2026](https://hai.stanford.edu/assets/files/ai_index_report_2026_chapter_2_technical.pdf)).

## What is available

- **Gemma 3**, released by Google on 12 March 2025, comes in four sizes. It was pre-trained on more than 140 languages, and the Gemma family had passed 100 million downloads ([Google](https://blog.google/technology/developers/gemma-3/)). Google says the largest version runs on one machine; its own chart suggests more hardware, so check before relying on it.
- **Meta MMS** (2023) does speech recognition and text-to-speech for more than 1,100 languages, but its licence does not allow commercial use ([InfoQ](https://www.infoq.com/news/2023/06/meta-mms-speech-ai/)).
- **Google WAXAL** (March 2026) released about 1,846 hours of transcribed speech in African languages under an open licence ([Google Research](https://research.google/blog/waxal-a-large-scale-open-resource-for-african-language-speech-technology/)). We could not confirm that Somali is included.

## Inside the country

Somali builders are already using open models. SOMA-S1 is a small Somali language model trained from scratch and released under an open licence ([Hugging Face](https://huggingface.co/boore1/SOMA-S1)). Jamhuriya University's SAIL lab released a Somali language model in March 2024 ([SAIL](https://www.sail.just.edu.so/news/sail-launches-somali-llm)) and in April 2026 launched a Somali model that detects false news ([Dawan](https://www.dawan.africa/news/jamhuriya-university-launches-somalias-first-somali-language-ai-lab-to-tackle-misinformation)). Open Somali speech models are published too ([Hugging Face](https://huggingface.co/Ogaal-Labs/Ogaal-ASR)).

## Why it is a tailwind

An open model can be owned and improved by Somalis, trained on Somali text and speech, and run without monthly bills to a foreign company or anyone's permission. Data can stay inside the country.

## What it means for young Somalis

A Somali university lab or a small start-up can now build tools in Somali, and in Maay, that match the quality of the world's best. The skills to do it are the skills young Somalis can learn now.

---
I want to help with this: email hello@calankii.com. Fix this page: email hello@calankii.com with the page link and your source.