This week: An Abu Dhabi AI lab released six open models that don't just chat in Arabic. They do the paperwork, too.

Elsewhere: How a model learns a new language without starting from scratch, plus how to run the small one on your own computer.

The New Prime Signal

Plenty of AI models can answer you in Arabic. Ask one to fix a spreadsheet while you explain the problem in Egyptian dialect, and things tend to get awkward. On October 15, the Institute of Foundation Models (IFM) at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) released Jais 3, a family of six open Arabic-centric models built to work as agents. The lab says it built them "for people and organizations across the Middle East and North Africa, and for Arabic speakers everywhere." Translation: less small talk, more getting things done.

What shipped: Six models, from 0.9B to 375B parameters, all built on K2 Horizon, IFM's own open model family. The 0.9B is meant for phones and edge devices, the 3.7B for laptops and local assistants, and the 7B for local and on-premise use. A 32B dense model and a 36B mixture-of-experts model (about 4B active) target single-server work, and the 375B flagship (23B active) needs a full 8-GPU node. Every size except the smallest gets a 512K-token context window (the 0.9B has 128K), which IFM says is room for a full contract or annual report in one prompt. The models handle Modern Standard Arabic, dialects such as Emirati and Egyptian, and Arabizi, which is Arabic written in Latin letters. And "open" here means more than weights: IFM also published checkpoints from every training stage, a sample of the training data, configs, training logs, and its evaluation code. Models and code are under Apache 2.0.

Why it matters: On IFM's agent tests, earlier Arabic-focused models such as ALLaM, Fanar, and Jais 2 "score close to zero." Jais 3 is trained to plan multi-step tasks, call tools and MCP servers, run terminal commands, search the web, and edit Arabic PDFs, PowerPoint decks, and Excel sheets, all while talking to the user in Arabic. On the lab's own benchmarks, the 7B beats Jais 2 70B, IFM's previous flagship, on Arabic (79.4 vs 66.7), and its agentic average beats Gemma 4 31B and Muse Glimmer 30B, open models more than four times its size. At the top end, the 375B posts the highest Arabic average of any open model IFM compared (84.0) and edges past Claude Sonnet 5 (82.4), though GPT-5.6 Sol and Gemini 3.8 Flash stay ahead. The usual caveat applies: these are the lab's numbers, run on its own test harness. To its credit, it cleaned up the tests first. On Arabic MMLU alone, IFM corrected 765 of its 14,455 questions.

What's next: Paired with Jais-ASR, IFM's speech recognition model, Jais 3 can take spoken instructions in Arabic or English and carry them out as an agent. The lab is asking builders to "try Jais 3 on your own tasks and tell us where it breaks," and it's pitching the recipe to everyone else: if you work on another language, the stage checkpoints and recipes let you adapt K2 Horizon the same way. Worth watching: whether other language communities take IFM up on that offer.

An AI that knows Emirati from Egyptian Arabic and can also update your spreadsheet is a welcome change from "works great, as long as you speak English." The models are the headline, but the open recipe might be the bigger gift. It's a how-to for giving any language its own capable assistant.

IFM didn't train Jais 3 from scratch. Each model starts from the finished K2 Horizon model of the same size, then goes through five more training stages. The lab says the switch to Arabic cost little of the foundation's general reasoning and English ability, and the Arabic gains were largest at the small sizes. The 3.7B, for example, scores 13 points ahead of Qwen3.5-4B on IFM's Arabic average.

How it works: First comes continued pre-training on 201.45B tokens, 75% Arabic and 25% "replay" of K2 Horizon's own training data, a mix IFM picked after also testing 90/10 and 50/50 splits. Next is a short cool-down on 3.47B tokens, then supervised fine-tuning on 41.1B tokens: about 22 million examples covering dialects, Arabizi, translation, grammar, poetry meter, agents, and tool use. Stage four fixes "doom loops," where a model's reasoning repeats the same span until it runs out of room. Averaged over the six models, the loop rate fell from 13.8% to 1.2%. Last is reinforcement learning on 19,769 prompts whose answers can be checked automatically, across instruction following, grammar, math, diacritics, and code. One surprising detail: about 59% of the Arabic training tokens are synthetic or machine-translated. That was on purpose, to add reasoning data and more of the Arabic people actually speak every day.

Try it: The 3.7B is the friendliest place to start. IFM's GGUF build on Hugging Face for llama.cpp comes in files from 3.16 GB (Q4_K_M) to 10.13 GB (BF16). K2 Horizon support is new in llama.cpp, so use a recent build, then run something like llama-server -hf IFM/Jais-3-3.7B-GGUF:Q4_K_M --jinja. The --jinja flag turns on the built-in chat template that handles reasoning and tool calls. On a Mac, the Jais 3 collection also has MLX versions. Start with a task you can check yourself, like summarizing an Arabic news article or turning a dialect message into Modern Standard Arabic. One heads-up from the model card: the quantized files haven't been separately evaluated yet, so a smaller download may not match the published scores.