Jais 3

IFM's Arabic-centric agent models, built on K2 Horizon in six sizes from 0.9B to 375B-A23B under Apache 2.0, with a 512K context.

In short. Jais 3 is a family of six open models that speak and work in Arabic: Modern Standard Arabic, dialects such as Emirati and Egyptian, and Arabizi. They are built on K2 Horizon, so they also plan tasks, call tools and edit Arabic PDFs, slide decks and spreadsheets. Sizes run from 0.9B for phones to 375B for a full GPU node.

Summary

Jais 3 is an Arabic-centric family from the Institute of Foundation Models (IFM) at MBZUAI, developed with Inception and Cerebras and announced on 15 October 2026. Its repositories first appeared on Hugging Face on 29 September. It has six sizes, 0.9B, 3.7B, 7B, 32B, 36B-A4B and 375B-A23B, and each starts from the final post-trained K2 Horizon model of the same size, so the architecture and the 512K-token context (128K for the 0.9B) carry over. IFM says it is the first adaptation of K2 Horizon to a new language and a fully sovereign UAE model, because every step from K2 Horizon's pretraining to the last reinforcement-learning stage was run at IFM. Training has five stages: continued pretraining on 201B tokens (75% Arabic and 25% replayed K2 Horizon data), a short cool-down, supervised fine-tuning on 41B tokens that covers Modern Standard Arabic, dialects, Arabizi, diacritics, poetry and agent trajectories, a preference step aimed at reasoning loops, and reinforcement learning with automatically checkable rewards. About 59% of the Arabic tokens are synthetic or machine-translated, and about 36% come from dialect-focused sources.

IFM ran the competing models itself on Arabic benchmarks that it audited and re-scored by generation rather than log-probability, so every number here is IFM's own. The 375B-A23B has an Arabic average of 84.0, against 82.0 for Qwen3.5-397B-A17B and 81.2 for Nemotron 3 Ultra; IFM puts it just ahead of Claude Sonnet 5 (82.4) and behind GPT-5.6 Sol (88.2) and Gemini 3.8 Flash (89.1). The 32B scores 83.1 and the 36B-A4B 82.5, against 82.6 for Gemma 4 31B, but Qwen3.8-27B is ahead of both on most agentic tests, such as Terminal-Bench 2.1 at 79.8 against 63.7 for the 32B, while scoring 77.4 on Arabic. The 7B scores 79.4 on Arabic against 76.9 for Gemma 4 12B and 70.9 for Qwen3.5-9B, and 70.2 on SWE-bench Verified against 50.8 for Qwen3.5-9B. The 3.7B and 0.9B score 78.1 and 68.5 against 65.4 for Qwen3.5-4B and 50.9 for Qwen3.5-2B. IFM also reports that its anti-loop preference step cut the average rate of endless reasoning loops from 13.8% to 1.2%.

With file tools, Jais 3 can read and edit Arabic PDFs, slide decks and spreadsheets, and IFM's Jais-ASR speech model, which has its own entry on this site, lets it take spoken instructions. BF16 weights are 758.3 GB for the 375B, 74.9 GB for the 36B-A4B, 69.6 GB for the 32B, 18.0 GB for the 7B, 10.1 GB for the 3.7B and 2.2 GB for the 0.9B, with an FP8 build of the 375B at 390.6 GB, and IFM also publishes FP8, MLX and GGUF builds for most sizes, such as the 7B at 5.6 GB in GGUF Q4_K_M. Hugging Face shows modest early traffic, about 2,100 to 3,600 downloads for each of the six main repositories in the 30 days to 11 November 2026. IFM publishes the checkpoints from every training stage, configs and logs, a sample of the training data and the audited Arabic benchmarks with their evaluation code, and the models and code are Apache 2.0.

Good for

  • Arabic-first assistants and agents that must handle Modern Standard Arabic, dialects such as Emirati and Egyptian, and Arabizi
  • Work on Arabic documents, such as contracts, decks and spreadsheets, using file tools and the 512K context
  • Local or on-device Arabic assistants with the 0.9B, 3.7B or 7B, or a single-server agent with the 36B-A4B
  • Teams adapting an open base model to another language, since the checkpoints from every stage and the recipes are published

Watch out for

  • Every score is IFM's own, on benchmarks it audited and re-scored itself, and Qwen3.8-27B beats the 32B on most agentic tests
  • The 375B needs an 8-GPU node in BF16, so most teams will use the smaller sizes or a hosted service
  • It is a text model, so speech input needs Jais-ASR or another recognizer in front of it
  • Early adoption is modest, with a few thousand downloads per repository, so community tooling and quantizations are still thin

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the 0.9B (BF16) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
0.9B (BF16) BF16 1.08B 2.48 GB 1.24 GB 0.745 GB Laptop or phone
3.7B (BF16) BF16 5.06B 12 GB 5.82 GB 3.49 GB Laptop or phone
7B (BF16) BF16 9B 21 GB 10 GB 6.21 GB Laptop or phone
7B GGUF Q4_K_M INT4 9B 21 GB 10 GB 6.16 GB Laptop or phone
32B (BF16) BF16 34.8B 80 GB 40 GB 24 GB Gaming GPU
MoVA-36B-A4B (BF16) BF16 37.4B 4B active 86 GB 43 GB 26 GB Gaming GPU
MoVA-36B-A4B GGUF Q4_K_M INT4 37.4B 4B active 86 GB 43 GB 25 GB Gaming GPU
375B-A23B (BF16) BF16 379B 23B active 872 GB 436 GB 262 GB Big-memory desktop
375B-A23B (FP8) FP8 379B 23B active 872 GB 436 GB 262 GB Big-memory desktop

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Jais 3?
Jais 3 is a family of six open models that speak and work in Arabic: Modern Standard Arabic, dialects such as Emirati and Egyptian, and Arabizi. They are built on K2 Horizon, so they also plan tasks, call tools and edit Arabic PDFs, slide decks and spreadsheets. Sizes run from 0.9B for phones to 375B for a full GPU node.
Can I use Jais 3 commercially?
Yes. Jais 3 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Jais 3 open source?
Jais 3 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Jais 3 on my own computer?
Yes. The smallest variant, 0.9B (BF16), needs about 0.745 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Jais 3 released?
Jais 3 was released by Institute of Foundation Models (MBZUAI) on October 15, 2026.
How long a context window does Jais 3 have?
Jais 3 supports a context window of up to 512K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
K2 Horizon Institute of Foundation Models (MBZUAI) 1.08B – 379B 23B active 512K Apache 2.0 (Permissive) 0.745 GB Laptop or phone
Jais-ASR Institute of Foundation Models (MBZUAI) 5.72B – Apache 2.0 (Permissive) 3.63 GB Laptop or phone
Qwen3.8-27B Alibaba Qwen 27.8B 256K Apache 2.0 (Permissive) 19 GB Gaming GPU
Gemma 4 Google DeepMind 5.1B – 30.7B 256K Apache 2.0 (Permissive) 3.52 GB Laptop or phone