Jais 3

// Overview

Jais 3 is a family of six open-weight language models built for Arabic, introduced on October 15, 2026 by the Institute of Foundation Models (IFM, part of the Mohamed bin Zayed University of Artificial Intelligence) and Inception42, G42's agentic AI company, in collaboration with Cerebras Systems. The six models run from 0.9B to 375B parameters: dense 0.9B, 3.7B, 7B and 32B models, a sparse Jais-3-MoVA-36B-A4B, and the flagship Jais-3-375B-A23B. They were released alongside Jais Speech, a separate speech recognition model for Arabic and its dialects. All weights are on Hugging Face under Apache-2.0, in Arabic and English.

The press release describes Jais 3 as the first adaptation of K2 Horizon, the six-model open fleet IFM released in September 2026, to another language. The models are trained on Modern Standard Arabic and regional dialects from the Gulf to North Africa, along with Arabic culture, poetry, cuisine and Islamic knowledge. On IFM's and Inception42's Arabic evaluations, the 7B model's average score rises from 56.4 for the K2 Horizon base to 79.6 for Jais 3. The model cards give a 512K-token context window for the 3.7B, 7B, 32B, 36B-A4B and 375B-A23B models and 128K for the 0.9B.

The flagship Jais-3-375B-A23B is a sparse Mixture-of-Experts model with 23B parameters active per token. Its model card compares it with Qwen3.5-397B-A17B, MiniMax M3, Qwen3.8-Flash-Next and NVIDIA Nemotron 3 Ultra: it posts the top scores on Arabic MMLU (89.3, tied with Qwen3.8-Flash-Next), Belebele (95.6), IslamicQA (90.2), dialect identification on QADI (71.1), FLORES200+ translation (91.9) and Arabic IFBench instruction following (93.9). On agentic tasks it scores 73.8 on Terminal Bench 2.1, 73.2 on GDPVal and 68.5 on MCPMark, the top MCPMark score in the table, while Qwen3.8-Flash-Next leads most other agentic rows. IFM notes that starred Arabic benchmarks were audited for sample quality and are scored on cleaned sets, published in its Arabic Benchmark Auditing collection.

The press release says the models can plan multi-step tasks, search the web, and read, edit and produce Arabic documents, spreadsheets, presentations and websites when connected to tools, and that the same models power the Jais chat and mobile apps. The 375B-A23B model card serves the model with vLLM or SGLang on 16 NVIDIA H100 GPUs across two nodes, or on a single 8-GPU H200 node, and recommends high reasoning effort for every request. Intermediate training checkpoints are published as branches of each repository, IFM publishes training logs on Weights & Biases, and GGUF, MLX and FP8 conversions sit in IFM's Hugging Face collections. The four smaller dense models also ship Uno diffusion adapters for lossless speculative decoding; the 0.9B adapter's card reports 1.51× faster generation on average with exactly the base model's output distribution.

// Benchmarks

This model's scores

1. Arabic MMLU (Jais-3-375B-A23B)89.3%
2. Belebele (Jais-3-375B-A23B)95.6%
3. IslamicQA (Jais-3-375B-A23B)90.2%
4. Arabic IFBench, prompt-level strict (Jais-3-375B-A23B)93.9%
5. Terminal Bench 2.1, Avg@3 (Jais-3-375B-A23B)73.8%
6. SWE-Bench Verified, offline (Jais-3-375B-A23B)70.2%
7. MCPMark, Pass@1 (Jais-3-375B-A23B)68.5%
8. Average Arabic score (Jais 3 7B, IFM and Inception42 evaluations)79.6

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• Six Apache-2.0 sizes from a 0.9B model small enough for a phone to a 375B-A23B flagship, sharing one training recipe
• 512K-token context on every size from 3.7B up (128K on the 0.9B)
• Jais-3-375B-A23B leads its model-card comparison on Arabic MMLU (89.3), Belebele (95.6), IslamicQA (90.2) and FLORES200+ translation (91.9)
• 93.9 on Arabic IFBench instruction following, against 81.7 for Qwen3.8-Flash-Next and 75.1 for Qwen3.5-397B-A17B
• Agentic tool use from the flagship: 68.5 on MCPMark, 73.8 on Terminal Bench 2.1 and 64.1 on Arabic File Processing
• Arabic benchmarks audited for sample quality and released, with training checkpoints and training logs published per model

// Best for

• Reach for it when you need open Arabic-English models you can self-host under Apache-2.0, for example for sovereign or on-premises deployments.
• Reach for the 375B-A23B flagship for Arabic agentic work: tool use, terminal tasks and producing Arabic documents and files.
• Reach for it for dialect-heavy Arabic text, translation into and out of Arabic, and instructions that must be followed exactly.
• Pick the 0.9B, 3.7B or 7B models for phones, laptops and edge hardware, using IFM's GGUF and MLX conversions.
• Look elsewhere for applications in languages other than Arabic and English, which the model cards list as out of scope.

// How to access