IFM and Inception42 release Jais 3 as six open-weight Arabic models

On 15 October 2026, the Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence and Inception42 published a Jais 3 press release. The page calls Inception42 G42’s sovereign agentic AI company and says the models were made in collaboration with Cerebras Systems. A WAM copy on Aletihad, timestamped 14:27 the same day with no timezone on the page, reports the same launch.

The release pairs six open language models, from 900 million to 375 billion parameters, with Jais Speech. The press release says they are trained on Modern Standard Arabic and regional dialects, and built on K2 Horizon, which it calls IFM’s fully open-source fleet launched last month. The six K2-Horizon base repositories under IFM list an Apache-2.0 license tag and created dates of 1 September 2026.

Six weight repositories

Six repositories under inception42 were public, each with safetensors files, and each card sets the license to Apache-2.0. A LICENSE text file was not at the root of the 375B or the 0.9B repository. Open weights means those public files plus the license line on the card.

The Hugging Face API lists createdAt of 29 September 2026 for the six, and lastModified of 15 October 2026 between 08:15 and 08:16 UTC.

Repository	What the card says
Jais-3-0.9B	Dense, 0.9B class, 128K context
Jais-3-3.7B	Dense, 3.7B class, 512K context
Jais-3-7B	Dense, 7B class, 512K context
Jais-3-32B	Dense, 32B class, 512K context
Jais-3-MoVA-36B-A4B	Mixture of experts, 36B stored, 4B per token, 512K
Jais-3-375B-A23B	Mixture of experts, 375B stored, 23B per token, 512K

The press release says all six share a 512,000-token window, “the longest of any Arabic-native model.” The 0.9B card says 128K. The other five cards say 512K.

Four Uno adapters are also public: Jais-3-0.9B-Uno, Jais-3-3.7B-Uno, Jais-3-7B-Uno, and Jais-3-32B-Uno. The 0.9B-Uno card describes a LoRA adapter that it says speeds up the frozen base and keeps the same output distribution. Those four cards do not set a license field.

The 375B card documents vLLM and SGLang serve commands, and the press release says the models are on Hugging Face with that support.

Speech is public. A set of 24 benchmarks is not.

Jais Speech is public as IFM/Jais-ASR, under IFM, not under inception42. The card sets Apache-2.0. The API lists createdAt of 6 October 2026, before the 15 October announcement. The card says the model covers Modern Standard Arabic and dialects from 19 Arab countries, code-switching into English, French, Spanish, and Italian, dialect identification, and optional diacritization, which is the vowel-mark feature in the press release. The press release names English and French for that switching. The card also says the model is based on Jais-3-4B. That base was not in the inception42 model list.

The press release says IFM and Inception42 tested the models on an audited suite of 24 Arabic benchmarks and are releasing the suite and the code to run it. The public collection Arabic-Benchmark-Auditing, updated 15 October 2026, held 16 items: 15 verified datasets and a discarded-samples set. Its description says defective items were removed by a three-judge audit. A public collection of 24 was not there. The 375B card says benchmarks marked with a caret will be released later.

Scores they report

Every number in this section is IFM and Inception42’s. The press release says the largest model matches the best open models on Arabic understanding and leads them on agentic tasks, and that the 7B and 32B models earned the highest average Arabic scores among open models tested at those sizes. It says the 3.7B model beat every competing open model of up to 5 billion parameters on Arabic and agentic tasks.

The Jais page gives Jais 3 375B an agentic average of 60.2, against 57.5 for MiniMax M3, 46.9 for Qwen3.5 397B, and 42.7 for Nemotron 3 Ultra. Closed models on that table sit higher. The 375B card’s SWE-bench Verified (Offline) row is 70.2 for Jais 3, 75.6 for MiniMax M3, and 80.8 for Qwen3.8-Flash-Next. The lead on agentic tasks is their summary of an average. Their own tables still place other models ahead on some rows.

The speech card reports a macro-average of 22.85% word error rate and 8.37% character error rate over six public test sets, and says that is the lowest of the models compared, including the commercial APIs from Speechmatics and ElevenLabs. It says Jais-ASR was measured with the leaderboard’s script, that the other open-weight scores were taken from the leaderboard, and that Speechmatics Melia 1 and ElevenLabs Scribe v2 were measured by IFM.

Haonan Li, a senior research scientist at IFM, says in the press release: “We built Jais 3 for Arabic as it’s really used.” The same page quotes Richard Morton on speech, writing, and agents. The WAM copy attributes a different Morton sentence, about the UAE’s investment in open AI research.

The six language-model cards, the Jais-ASR weights, and 15 verified audit sets are public. The figure of 24 benchmarks remains the organizations’ count of what they tested.