Jais 3 for Arabic Agents: Weights, Benchmarks and Hosting Trade-offs

Jais 3 is a family of open-weight Arabic-native models released on 15 October 2026 by MBZUAI’s Institute of Foundation Models (IFM) and Inception42. Every published evaluation was run by the developers, and no independent test was found. Use the tables to pick candidates, then test them yourself.

The short answer

Sizes: six base sizes: 0.9B, 3.7B, 7B, 32B, MoVA-36B-A4B (4B active) and 375B-A23B (23B active).
License: Apache-2.0 on the 7B, 32B, MoVA-36B and 375B cards. The technical report and code are “Coming soon” on the 32B and 375B cards, while weights are out with CPT, cooldown, FTPO and SFT checkpoint branches.
Context: 512K tokens, on the same four cards.
Lineage: built on K2 Horizon, IFM’s open model family from September 2026 (Gulf News).
Serving: a standard vLLM or SGLang path.
Evidence: developer-run only. Jais leads on some Arabic tests and trails Gemma on others, and the card claims no overall win.

What is in the family?

The Hugging Face collection lists ten repositories under the inception42 organization: the six base sizes plus an “Uno” variant of the first four. No card that was opened defines Uno.

The 7B and 32B are dense decoder-only models. MoVA-36B-A4B is a mixture-of-experts with “Mixture-of-Values attention”, and 375B-A23B is a sparse MoE. The IFM page also shows MLX builds in bf16, 8-bit and 4-bit, and the 32B card says FP8, GGUF and MLX builds exist.

The active-parameter figures matter for cost. The 375B model activates 23B parameters per token, and MoVA-36B activates 4B. Compute per token follows the active part, but you still need memory for all the weights.

How do you serve it?

The 32B card documents vLLM 0.30.0 or later or SGLang 0.5.20 or later, BF16 weights, tensor parallelism of 2, and a k2_horizon reasoning parser.

Launch coverage says the 7B Arabic average rose from 56.4 for the K2 Horizon base to 79.6 for Jais 3 (Gulf News, Middle East AI News, TechEdgeAI). By our own subtraction that is a gain of 23.2 points. The 7B card itself has no Arabic-average row, so treat it as a reported figure only. No training data release is mentioned in the sources.

What does the 32B card report?

These are developer figures. The card does not say who ran the evaluations (the citation credits IFM, Inception42 and Cerebras). The comparators are open models: Gemma-4-31B-it and Qwen3.8-27B.

Benchmark	Jais 3 32B	Gemma-4-31B-it	Qwen3.8-27B
Arabic IFBench (instruction following)	88.7	80.8	81.3
Arabic IFEval	80.0	78.7	77.0
SimpleQA Verified (Ar), aggregate	28.1	10.6	8.0
SimpleQA non-hallucination rate	70.1	10.5	5.4
HalluScore	67.3	69.7	54.6
AraGen v3	71.0	73.4	66.7
Arabic MMLU	88.6	89.5	85.8
Belebele	95.0	95.9	94.4
Alyah	83.0	88.4	77.7
FLORES200+	91.5	92.8	81.8
QIAS-1	95.9	93.2	81.6
BrowseComp-Plus	71.9	42.5	53.0
MCPMark	66.1	48.8	49.6
Terminal Bench 2.1	63.7	43.5	79.8
SWE-Bench Verified	72.4	59.0	73.5

The table splits three ways.

Jais leads on: Arabic instruction following, SimpleQA, QIAS-1, BrowseComp-Plus and MCPMark.
Gemma leads on: Arabic MMLU, AraGen, Belebele, Alyah, FLORES and HalluScore.
Qwen leads on: Terminal Bench and, narrowly, SWE-Bench.

The standout is the SimpleQA non-hallucination rate: 70.1 for Jais against 10.5 and 5.4. On Arabic IFBench, our subtraction gives a lead of 7.9 points over Gemma (88.7 minus 80.8) and 7.4 over Qwen (88.7 minus 81.3). Gemma’s edge on Arabic MMLU is under one point.

The card claims no overall win. Launch coverage says the 32B has the “highest average Arabic score among open models at its size”. The card has no Arabic-average row, so the “83.1 versus 82.6” figures that circulate in summaries are not reproduced here.

What is the “audited” benchmark package?

IFM publishes “Arabic Benchmark Auditing”, 16 datasets in which a three-judge LLM audit removed defective items. The discarded samples and verdicts are released.

The audit covers only parts of some datasets. The dialectal Arabic MMLU audit covers dialect rows only (935 of 15,675 removed), and MedAraBench has the highest removal rate (896 of 4,927, or 18.2%). Starred benchmarks on the cards are the audited ones; Belebele, FLORES200+ and AraGen are not. This is the developer auditing its own benchmarks, not an outside review.

What is Jais Speech?

Jais-ASR (about 6B parameters, Apache 2.0) is built on Jais-3. It covers Modern Standard Arabic plus dialects from 19 Arab countries, with code-switching into English, French, Spanish and Italian. A vLLM plugin exposes an OpenAI-style transcription endpoint.

On the Open Universal Arabic ASR Leaderboard it scores 22.85% WER and 8.37% CER averaged over six test sets. The closest open rival, Audar-ASR-V1-Turbo, scores 23.17% and 9.23%. It does not lead everywhere: on SADA it scores 30.85% WER against 33.36% for Voxtral-Small-24B. Scores for commercial services such as Speechmatics and ElevenLabs were measured by IFM itself.

Who should use which model, and how do you test it?

Pick by the rows that match your product. Instruction following and factual QA favored Jais; broad knowledge rows favored Gemma on this card.

Build 20 real tasks in Arabic, in your users’ dialect, and run them against two candidates with the same settings.
Match benchmark rows to your use before you trust a headline.
Plan hardware first. The 32B model is documented with tensor parallelism of 2, meaning at least two GPUs.

To measure what matters after that, see why a cheaper model can cost more per finished task.

Questions and answers

Is Jais 3 the best Arabic model?

Not on this evidence. The 32B card shows it ahead on instruction following and some factual tests and behind Gemma on most knowledge tests, and all evaluations are developer-run.

What license does it use?

Apache-2.0 on the model cards that were opened. The technical report and code are listed as coming soon.

How big is the largest model?

375B-A23B, a sparse mixture-of-experts with 23B active parameters.

Does it handle the Saudi dialect?

Jais Speech claims coverage of dialects from 19 Arab countries. Test it on recordings of your own users before relying on that claim.

How do I run the 32B model?

Use vLLM 0.30.0 or later or SGLang 0.5.20 or later, with BF16 weights and tensor parallelism of 2.

Terms in plain language

Open weights: the trained model files can be downloaded and run on your own hardware.
Mixture-of-experts (MoE): a model that uses only part of its parameters for each token. “Active” means the part used.
WER and CER: word and character error rates in speech recognition. Lower is better.
Tensor parallelism: splitting one model across several GPUs.

Sources

Facts last checked against the linked pages on 18 October 2026. Benchmark figures are developer-reported.