Jais 3 Launches: Six Free Open Arabic Models and a Dialect-Aware Speech Model

Abu Dhabi’s MBZUAI Institute of Foundation Models (IFM) and Inception42 released Jais 3 on 15 October 2026: six open-weight language models, from 0.9 billion to 375 billion parameters, plus a speech-to-text model built for spoken Arabic. The practical headline is the licence. Every model card we checked on Hugging Face says Apache 2.0, so you can download, run, fine-tune and ship these models commercially without paying anyone.

What actually shipped

We went to the model repositories rather than relying on the press release. The family breaks down like this:

• Four dense models at 0.9B, 3.7B, 7B and 32B. These are the ones most people can run on their own hardware.
• Two mixture-of-experts models: a 36B model that activates about 4B parameters per token, and the 375B flagship, which activates about 23B.
• Jais-ASR (branded “Jais Speech” in the announcement), a transcription model covering Modern Standard Arabic and dialects from 19 countries.

The cards list a 512K-token context window and say the models are built on K2 Horizon, the open model fleet IFM released in September. The team also published ready-made quantised builds: GGUF for llama.cpp, MLX for Apple silicon Macs, and FP8 for servers.

Why it matters if you work in Arabic

Most frontier models handle written Modern Standard Arabic well and stumble on the dialects people actually speak. Jais 3 is aimed squarely at that gap. Gulf News reports that the models were trained on MSA alongside Emirati, Gulf and North African dialects. The speech model is the more interesting piece. According to its model card, it keeps English, French, Spanish or Italian words in their own script when a speaker code-switches, which describes most business meetings from Dubai to Casablanca. It can also identify the dialect, add diacritics on request, and say there is no speech instead of hallucinating text from a silent recording.

We are not reprinting benchmark numbers. They come from the developers, and there is still no neutral Arabic benchmark we trust. One detail is worth knowing, though. IFM’s own comparison table for the 7B model shows a larger Google model ahead of it on open-ended Arabic generation. “World-leading” holds only on some tests.

What it costs, and who should care

The weights are free. Neither the announcement nor the model cards mention a consumer chat app or a priced hosted API, so your real cost is the hardware you run it on. Going by the published file sizes, the 4-bit (Q4_K_M) builds are roughly 0.7 GB for the 0.9B model, 3.2 GB for 3.7B and 5.6 GB for 7B. That is laptop territory. The 375B flagship is for enterprises and government bodies that want data kept on their own servers, and the announcement addresses that group directly.

Three groups gain the most. The first is developers building Arabic products who would rather not send customer data abroad. The second is journalists and creators transcribing dialect interviews. The third is public-sector teams with data-residency rules.

What to do now

1. Start small. Pull the 3.7B or 7B GGUF from the Jais 3 collection on Hugging Face.
2. Update your runtime first. The card warns that support for the K2 Horizon architecture has only just been merged into llama.cpp. Older builds of llama.cpp, and of tools built on it such as Ollama and LM Studio, may fail to load it.
3. Run your own dialect test. Record two minutes of real conversation in your dialect. Transcribe it with Jais-ASR and with whatever tool you use today, then count the wrong words in each transcript. That tells you more than any leaderboard.
4. For writing, ask the 7B model to draft a formal email and then rewrite it in your dialect. Put the same request to ChatGPT and compare the two answers.

For background, our earlier piece Does a model that speaks Arabic really understand it? sets out the questions worth asking. With open weights, Jais 3 lets you answer them yourself instead of taking anyone’s word for it.

Sources: IFM and inception42 model cards on Hugging Face, Gulf News, and the press release via Zawya. IFM’s own release page blocks automated requests, so we could not read it directly.