MBZUAI and Inception42 open Jais 3, six Arabic models plus speech under Apache-2.0

What did MBZUAI and Inception42 release?

Arabic-language AI has long leaned on English-first models that are adapted after the fact. On 15 October 2026, Middle East AI News reported that Mohamed bin Zayed University of Artificial Intelligence’s Institute of Foundation Models (IFM) and Inception42, G42’s sovereign agentic AI company, launched Jais 3: six open language models plus a speech recognition model aimed at dialects, writing and multi-step agents.

IFM’s own Jais 3 product page frames the release as Arabic-native rather than translated: the models are built on K2 Horizon, the fully open foundation fleet IFM announced in September 2026, and are meant to serve more than 400 million Arabic speakers with culturally grounded behaviour as well as linguistic accuracy.

The same page links a press release and separate technical blogs for the LLM and ASR stacks. We grounded this article on the product page, the Hugging Face model cards for the 375B flagship and Jais-ASR, the Middle East AI News write-up, and IFM’s earlier K2 Horizon press release for the base fleet’s licensing posture.

How big is the Jais 3 fleet, and what can each size do?

The language fleet spans six sizes that share one training recipe, agent capabilities and—according to IFM—a 512,000-token context window that the developers call the longest of any Arabic-native model. Middle East AI News lists the range as 900 million to 375 billion parameters; Hugging Face hosts matching inception42 and IFM repositories for 0.9B, 3.7B, 7B, 32B, MoVA 36B-A4B and 375B-A23B checkpoints, including FP8, GGUF and MLX builds for smaller sizes.

The flagship Jais-3-375B-A23B model card describes a sparse mixture-of-experts model that stores 375 billion parameters and activates about 23 billion per token, with a 512K context window. An FP8 build is published under the IFM org. The card’s reference serving recipe targets 16× NVIDIA H100 GPUs across two nodes for vLLM, or a single 8-GPU H200 node when memory allows.

That architecture sits in the same open-MoE conversation as other large sparse releases this week, including StepFun’s Step 5 Preview. The practical difference for Arabic builders is the shared post-training: IFM says every Jais 3 size inherits the same agent workflows for documents, slides, spreadsheets and web pages, so teams can prototype on a phone-class checkpoint and keep prompts when they move up.

Middle East AI News adds that each model ships with five training checkpoints for researchers, and that UAE and international government entities can adapt and host the models under sovereign AI constraints. Cerebras Systems is named as a collaboration partner on the family.

What do the developers’ own Arabic and agent scores say?

IFM publishes a large Arabic and agentic scoreboard on the Jais 3 page and a detailed table on the 375B model card. Those figures are the developers’ own. On the product page’s small-model table, Jais 3 3.7B posts an Arabic Average of 78.1 and an Agentic Average of 44.5, ahead of the listed open comparators under about 5 billion parameters. The mid-size table puts Jais 3 32B at an Arabic Average of 83.1 and an Agentic Average of 60.0.

For the flagship, the page’s frontier table shows Jais 3 375B with an Arabic Average of 84.0 and an Agentic Average of 60.2, ahead of the open models shown on Arabic Average and competitive with closed frontier systems that IFM lists only as reference. The 375B card expands that into audited Arabic MMLU, dialect, Islamic knowledge, instruction-following and agent suites such as Terminal Bench 2.1, SWE-Bench Verified and Arabic File Processing.

Middle East AI News highlights a before-and-after claim at 7 billion parameters: adapting K2 Horizon to Arabic raised the average Arabic score from 56.4 on the base model to 79.6 on Jais 3. That is a vendor measurement, but it is the clearest public statement of how far language-and-culture adaptation moved the K2 Horizon base without training a new family from scratch.

IFM also says starred benchmarks were audited for sample quality, with ambiguous or invalid items removed, and that the audit materials live in an IFM Hugging Face collection. We have not reproduced the scores or the audit.

How does Jais Speech fit the stack?

Jais Speech, published as IFM/Jais-ASR under Apache-2.0, is the speech half of the stack. The model card tags Modern Standard Arabic and a long list of dialect codes, plus English, French, Spanish and Italian for code-switching. IFM says the system keeps Latin-script words in their original script, identifies the dialect and can add Arabic vowel marks.

On IFM’s product-page table drawn from the Open Universal Arabic ASR Leaderboard, Jais Speech records the lowest average word and character error rates among the models shown (22.85 / 8.37). Demo clips on the page compare dialectal Saudi, Moroccan and Egyptian speech against commercial ASR systems; those clips and error rates are IFM’s own exhibits.

Paired with the language models, the ASR is meant to turn spoken Arabic or mixed Arabic-English instructions into the same agent workflows as text. That is a different problem from Emirati-focused ASR packages such as TII’s Falcon Emirati ASR and OCR launch: Jais Speech is pitched as a multi-dialect open stack that feeds a full agent platform rather than a single national variant.

How do you get the weights, and what does Apache-2.0 allow?

The inspected Hugging Face cards for Jais-3-375B-A23B and Jais-ASR declare Apache-2.0. That is a permissive open-weights versus open-source posture: you can use, modify and redistribute under the licence terms without a separate commercial rider on the cards we read. Always re-check the licence file in the exact revision you download.

Serving instructions on the 375B card cover vLLM (≥0.30.0) and SGLang (≥0.5.20) with K2 Horizon reasoning and tool-call parsers. Recommended sampling differs for English and Arabic thinking modes; the card asks for at least 32,768 output tokens when reasoning is enabled. Smaller quantised builds under the IFM org target local and on-device experiments.

IFM’s September K2 Horizon press release said that base fleet also shipped weights, code, training data and methodologies under Apache-2.0. Jais 3 is presented as an Arabic specialisation of that fleet rather than a closed fork, which matters for labs that want to audit how dialect and culture adaptation changed the base.

Why Arabic-native open models matter now

Most frontier assistants still treat Arabic as a secondary language. A six-size open fleet with shared agents and a dedicated ASR model gives governments, banks and startups in the Gulf and North Africa a path to host Arabic workflows without sending every token to a US closed API.

It also continues IFM’s open-science pitch after K2 Horizon: publish the adaptation so others can measure it. Whether independent labs will confirm the Arabic Average and agent numbers is an open question—the same caveat that applies to other vendor scoreboards this month. For builders, the immediate product is clearer: downloadable checkpoints, Apache-2.0 terms on the cards we inspected, and recipes that match the mixture-of-experts serving stack many teams already run.

For multilingual product teams outside the region, Jais 3 is also a test of whether language-first post-training on a fully open base can beat larger English-centric models on dialect, culture and Arabic file work. IFM’s own tables say yes at several sizes; buyers should treat that as a claim to verify on their own documents and call recordings.

What is still unverified

We have not downloaded multi-hundred-gigabyte weights, run Arabic or agent benchmarks, or tested Jais Speech on our own audio. The 15 October launch date comes from Middle East AI News and the product framing on ifm.ai; some Hugging Face repositories show earlier createdAt timestamps for quantised builds, so treat the public launch narrative and the upload chronology as related but not identical.

IFM’s press-release and technical-blog URLs linked from the product page were not fully retrieved as static HTML in this run (the marketing site is heavily client-rendered). Pricing for any hosted Jais agent product, enterprise support terms, and independent third-party leaderboard confirmations are not established here. Trademark and brand-asset reuse for Jais remain separate from the Apache-2.0 model licence—hence the AI-generated cover rather than a vendor mark.