Jais 3 ships six Arabic-native open models, from phone-size to 375B
On October 15, 2026, the Institute of Foundation Models (MBZUAI) and Inception42 released Jais 3, a fleet of six open models built for Arabic — from 0.9B to 375B parameters — paired with a speech recognition model for the region’s dialects. For organizations that work in Arabic, it is now the sovereign baseline to test before any Western general-purpose model.
October 15, 2026. The Institute of Foundation Models (IFM), part of Abu Dhabi’s MBZUAI, and Inception42 release Jais 3. Six models. The fleet runs from a 0.9-billion-parameter model, light enough to run on a phone, up to a 375-billion flagship. 512,000 tokens. That is the advertised context window, the longest of any Arabic-native model. Why it matters: for the 400 million Arabic speakers and the hospitals, administrations and companies that work in the language, the question is no longer “can a general-purpose model speak Arabic,” but “does a model trained for Arabic from the start do better” — and Jais 3’s answer is quantified.
A fleet of six models, from phone to flagship
Jais 3 is the first adaptation of K2 Horizon — the fleet of six fully open models IFM released in September — to a new language. The mechanics matter: rather than retraining a Western LLM, you start from an open foundation and adapt it to Arabic with the same training recipe. All six models share that recipe, their agentic capabilities, and a 512,000-token context window.
The scale is deliberately graduated. The 0.9-billion model fits on a mobile device, the 3.7-billion targets edge and small servers, the 7-billion the workstation, the 32-billion the enterprise server, and the 375-billion cluster deployments. Each organization picks the size that matches its hardware and budget — decisive for administrations that mandate on-premises deployment for sovereignty reasons. Every model ships with five training checkpoints, letting a researcher trace its development or resume from any stage.
IFM gives the adaptation proof in one number: on its Arabic evaluations, the 7B model’s average score rises from 56.4 for the K2 Horizon base to 79.6 for Jais 3. An open foundation can carry over to a new language and culture without starting from scratch.
Instruction following and real work
General-purpose models have a known Arabic weakness: they drift between dialects and registers within a single document. Jais 3 claims the opposite — fluent, consistent writing and strict adherence to complex instructions. The press release’s example is telling: given four instructions for a hospital appointment letter, including a word to avoid, Jais 3 met all four, while a general-purpose model used the banned word twice.
The model is also tooled, not merely textual. Connected to software tools, Jais 3 plans multi-step tasks, searches the web, and reads, edits and produces Arabic documents, spreadsheets, presentations and websites. The release’s example — a four-slide deck on the El Niño 2026–2027 forecast, built from published data after research — illustrates full agentic use rather than simple completion. The same models power the Jais chat app, and an agentic platform for everyday workflows is coming.
The openness strategy deserves emphasis. IFM publishes not only the weights but also the 24 audited Arabic benchmarks used for evaluation and the code to run them. That is what separates a marketing claim from a verifiable measurement: a third party can reproduce the numbers on its own data.
The tool use is where Arabic-specific value compounds. A Western model that handles Arabic documents passably still struggles with right-to-left layout, mixed-script tables and the formal register of administrative prose. Jais 3’s examples — a hospital appointment letter that respects every constraint, a four-slide deck assembled from live data — are precisely the tasks a MENA organization cannot hand to a generalist without a round of manual cleanup.
Jais Speech: real Arabic, dialects and code-switching included
The most technical piece is Jais Speech, the paired speech recognition model. It transcribes Modern Standard Arabic and the dialects spoken in 19 Arab countries, including in noisy, real-world conditions. Its distinguishing feature is code-switching: it follows speakers who slip from Arabic into English or French mid-sentence and back, keeping those words in their original script. It also identifies the dialect being spoken and can add vowel marks to a transcript.
Across six public test sets, Jais Speech posts the lowest average word and character error rates of the models compared, including commercial speech services. That is the kind of result that never shows up in a stage demo but decides whether a transcriber is actually useful in a call center, a clinic or a newsroom. Paired with the language models, it lets Jais 3 take spoken instructions in Arabic or English and act on them.
“Building an Arabic LLM is not a translation exercise”
The launch’s most important sentence comes from Ashish Koshy, CEO of Inception42: “Building an Arabic LLM is not a translation exercise. It means training on the dialects people actually speak, the way they write, and the culture the language carries.” It sums up the project’s thesis, and it is verifiable in the technical choices: training on Modern Standard Arabic and the dialects from the Gulf to North Africa, incorporating culture — poetry, cuisine, Islamic knowledge — and handling mid-sentence language switches.
The positioning is resolutely sovereign. Jais 3 is built by IFM at MBZUAI and Inception42, G42’s sovereign agentic AI company, in collaboration with Cerebras Systems. Enterprises and governments can adapt the models and choose where to deploy them, in line with data-sovereignty requirements — a decisive argument in a region where regulators push for local hosting.
The buying question is easy to frame. A general-purpose model that “also speaks Arabic” treats Arabic as an interface option, not a working language. Jais 3 inverts the priority: Arabic is the native language, English the second. For a hospital writing appointment letters, a ministry producing regulatory documents, or a newsroom transcribing interviews, that difference is operational before it is symbolic.
Verdict
If your organization works primarily in Arabic — healthcare, government, education, media across the MENA region — Jais 3 is now the default entry point: the weights are open, the audited benchmarks are published, and the models deploy on your own infrastructure. Match size to hardware: the 3.7B or 7B for edge and workstations, the 32B or 375B for heavy workloads and agents.
If you build Arabic software on mobile or at the edge, the 0.9B and 3.7B are the natural candidates — the 3.7B outperforms, per IFM, every competing open model up to 5 billion parameters on Arabic and agentic tasks. If your need is transcribing real Arabic speech, test Jais Speech first: its public-benchmark results against commercial services justify an evaluation on their own.
As with any open model, validate on your own distributions before generalizing. But for once, the audited benchmark and the code ship with it — which turns that validation from an act of faith into a reproducible procedure.