Jais

Jais 3 is a range, not a single model

Jais 3 was announced on 15 October 2026 (our news report: Jais 3 launches: six free open Arabic models). It comes in four dense sizes, 0.9B, 3.7B, 7B and 32B parameters, plus two mixture-of-experts models: a 36B model that activates about 4B parameters per token, and a 375B flagship that activates about 23B.

Every main model card lists Apache 2.0. The 7B card describes a 512K-token context window, adjustable reasoning depth and tool calling, so these are meant for agents and applications, not just chat. The reasoning has a cost you should plan for: the card recommends high reasoning effort and at least 32,768 output tokens, so answers on a laptop will take a while.

The previous generation, Jais 2 (December 2025, 8B and 70B chat models), is still on Hugging Face but now requires you to sign in and accept terms before downloading. Unless you already built on it, there is little reason to pick it over Jais 3.

Three ways to try it

Checked in October 2026 on the Hugging Face model pages and jaischat.ai. The chat site served us only its loading shell, so we cannot confirm its sign-up rules. The team also publishes MLX builds for Apple silicon Macs and FP8 builds for servers.

One practical warning. Jais 3 uses a new architecture, K2 Horizon, whose support only recently landed in llama.cpp. Older versions of LM Studio and Ollama may refuse to load it, so update first. Jais is not in Ollama’s official library yet; the copies you will find there were uploaded by community members, not the developers. If running models locally is new to you, read How to Run AI Models on Your Own Computer and then Ollama vs LM Studio.

The real test is dialect, and Jais 3 brings a speech model to it

Most frontier models write good Modern Standard Arabic and stumble on the dialects people actually speak. Jais 3 adds Jais-ASR, a speech-to-text model whose card claims coverage of MSA and dialects from 19 Arab countries. It keeps English, French, Spanish and Italian words in their own script when a speaker switches languages mid-sentence, can add Arabic diacritics on request, and labels the dialect it hears. That is a fair description of an ordinary business meeting anywhere from Dubai to Casablanca.

We do not reprint benchmark scores, which come from the developers themselves. A five-minute test tells you more. Ask the 7B model a question in your own dialect that mixes in an English technical term, then put the same question to ChatGPT and judge which answer sounds like a real person from your region. If you have a recording of dialect speech, transcribe it with Jais-ASR and with your current tool, and count the wrong words in each.

Who should use it, and who should look elsewhere

Jais fits developers building Arabic products who would rather not send customer data abroad, organisations in the Gulf with data-residency rules, and journalists or researchers transcribing dialect interviews. The card is clear about its limits: it does not support languages other than Arabic and English, and it should not drive medical, legal or financial decisions without human review.

If you want a ready-made Arabic assistant with images and voice and no setup at all, try Qatar’s Fanar. In Saudi Arabia, compare it with ALLaM. For how these projects differ and when each makes sense, see our guide Arabic AI Models: Does a Model That Speaks Arabic Really Understand It? Official links: Jais Chat and the model pages on Hugging Face.