Jais 3: six open Arabic-native models — from a 900M phone runner to a 375B flagship, with speech that follows your code-switching

The Institute of Foundation Models at MBZUAI and Inception42 introduced Jais 3 on Thursday: six open language models built natively for Arabic — dialects, code-switching and all — with a 512,000-token context window, plus a speech model that claims the lowest error rates of any open Arabic recognizer.

A fleet, not a model

The Institute of Foundation Models (part of Mohamed bin Zayed University of Artificial Intelligence) and Inception42 introduced Jais 3 on Thursday — a fleet of six open language models built for Arabic, from a 900-million-parameter model small enough to run on a phone to a 375-billion-parameter flagship. Alongside the language models comes Jais Speech, a speech recognition model covering Arabic and its regional dialects. Everything is built on top of K2 Horizon, IFM's fleet of six fully open-source foundation models released last month.

Jais 3 is trained to understand Modern Standard Arabic and the dialects people actually speak, from the Gulf to North Africa — plus the culture around the language, including its poetry, cuisine, and Islamic knowledge. For the more than 400 million people who speak Arabic, that decides whether an AI tool can be trusted with everyday work, the lab argues. Every model shares the same training recipe and agent capabilities, and each gets a 512,000-token context window — the longest of any Arabic-native model, per IFM.

"Not a translation exercise"

The loudest idea in the launch is that native beats translated. "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," said Ashish Koshy, CEO of Inception42. "We built Jais 3 for Arabic as it's really used," added Haonan Li, a senior research scientist at IFM — including the habit of slipping into English or French mid-sentence and back again, which Jais 3 handles without forcing everything into one script.

On IFM's and Inception42's audited suite of 24 Arabic benchmarks, the largest Jais 3 model matches the best open models on Arabic understanding and leads them on agentic tasks. The advantage holds down the ladder: the 3.7B model outperformed every competing open model of up to 5 billion parameters on Arabic and agentic tasks, and the 7B and 32B models earned the highest average Arabic scores among open models at their sizes. At 7B, the average score across IFM's and Inception42's Arabic evaluations jumps from 56.4 for the K2 Horizon base model to 79.6 for Jais 3 — the lab's exhibit A that an open foundation can be carried into a new language and culture without starting from scratch.

IFM's own published comparison table is blunt: Jais 3 3.7B averages 78.1 on Arabic benchmarks against 65.4 for Qwen3.5 4B and 64.5 for Gemma 4 E2B, and scores 68.0 on SWE-bench Verified where Qwen3.5 4B manages 44.0. These are vendor-reported numbers on vendor-designed Arabic benchmarks, so independent verification will matter — but IFM is releasing the audited benchmarks and the code to run them, and each model ships with five training checkpoints so researchers can trace its development.

Speech that follows the conversation

Jais Speech may be the more practical half of the launch. It transcribes Modern Standard Arabic and dialects spoken across 19 Arab countries in noisy, real-world conditions, follows speakers as they switch between Arabic and languages like English or French — keeping those words in their original script — identifies the dialect being spoken, and can add Arabic vowel marks to a transcript. Across six public test sets, IFM and Inception42 say Jais Speech recorded the lowest average word and character error rates of the models compared, including commercial speech recognition services. Paired with the language models, it lets Jais 3 take spoken instructions in Arabic or English and act on them.

The agentic pitch

The third leg is work. Connected to software tools, Jais 3 can plan multi-step tasks, search the web, and read, edit, and produce Arabic documents, spreadsheets, presentations, and websites. The launch demo: asked for an Arabic presentation on the 2026–2027 El Niño forecast, it researched the topic and delivered a four-slide deck built from published forecast data. "Stronger reasoning and instruction following, combined with tool use, provide a foundation for multi-step agentic workflows," said Dr. Larry Murray, SVP of Applied Science at Inception42. The same models power the Jais chat app and mobile app, and a Jais agentic platform for everyday workflows is coming soon.

The instruction-following claim got a vivid test: given four instructions for an Arabic hospital appointment letter — including one word to avoid — Jais 3 met all four, while a general-purpose model used the banned word twice. "Jais 3 brings deep Arabic understanding to speech, writing and agents," said Richard Morton, vice president and executive director of IFM. "It is built for the way Arabic is spoken, the way it is written, and the work people need done in it."

What it means

Jais 3 is a bid for the whole Arabic AI stack — speech in, writing out, agents that do the work — and a proof point for a strategy: build one strong open foundation (K2 Horizon), then adapt it per language instead of per model family. If the adaptation story holds, the marginal cost of a genuinely native model for another language falls sharply, and the countries that spent years watching English-first models treat their language as an afterthought stop having to choose between sovereignty and quality. The benchmark claims deserve the usual skepticism until independent labs rerun them — the audited, open benchmark suite is the launch's most checkable promise. Jais 3 and Jais Speech are available today on Hugging Face, with deployment support for vLLM and SGLang, in collaboration with Cerebras Systems.

Sources: Institute of Foundation Models press release: Jais 3 launch (Oct 15, 2026); IFM: Jais 3 technical page with benchmarks.