Run Jais 3 locally: GGUF, MLX and Mac memory by size

Short answer

Yes, you can run Jais 3 locally. IFM ships official GGUF files for llama.cpp and MLX builds for Apple silicon at every size up to the 36B MoVA model. The 7B at Q4_K_M is a 5.6 GB file and suits a 16 GB Mac. The 32B at Q4_K_M is 21.1 GB and wants 48 GB. The 375B flagship is not a Mac model. There is no official Ollama library tag yet, and I have not run any of these files myself.

File sizes by Mac ↓How to run it ↓Model-fit calculator →

What Jais 3 is

Jais 3 is six open language models, from 0.9B to 375B parameters, plus a separate speech model called Jais Speech. IFM built them on K2 Horizon, its own open model family from September, and trained them for Modern Standard Arabic, dialects such as Emirati and Egyptian, and Arabizi. All six are tagged Apache 2.0 on Hugging Face. The context window is 512K tokens, except the 0.9B, which has 128K.

The sizes are 0.9B, 3.7B, 7B and 32B dense models, a 36B mixture-of-experts model with about 4B active per token (MoVA-36B-A4B), and the 375B-A23B flagship. IFM also publishes checkpoints from each training stage, which is rare for a release this size.

I work from Dubai, so an Arabic model that runs on a laptop matters to me: a small UAE team could keep Arabic documents off third-party servers. Below is which Jais 3 file fits which Mac.

Jais 3 file sizes and the Mac each one needs

These are IFM’s own GGUF and MLX files, read from the Hugging Face file lists on 4 November 2026. The “Mac” column is my planning rule, not a measurement: keep the weights near half of unified memory, so there is room for context, other apps and macOS. The model-fit calculator does the same sum with your own context length.

Jais 3 official local builds (vendor file sizes, GB)
Model	GGUF Q4_K_M	GGUF Q8_0	MLX 4-bit	Mac I would plan for
Jais-3-0.9B	0.67	1.15	0.86	Any Mac
Jais-3-3.7B	3.16	5.39	4.2	8 to 16 GB
Jais-3-7B	5.59	9.57	7.2	16 GB (Q4), 24 GB (Q8)
Jais-3-32B	21.08	36.97	25.5	48 GB (Q4), 64 GB+ (Q8)
Jais-3-MoVA-36B-A4B	22.37	39.83	26.4	48 GB (Q4), 64 GB+ (Q8)
Jais-3-375B-A23B	None	None	None	Not a Mac model (about 758 GB in BF16; IFM ships FP8 for servers)

MLX figures are repo totals, which is why some are larger than the GGUF of the same bit width. Every size up to 36B also has 8-bit and BF16 MLX builds. On a 48 GB Mac I would look at MoVA first. It needs the memory of a 36B model but only computes about 4B parameters per token, so it should decode much faster than the dense 32B. Nobody I could find has published a Mac speed for it yet.

How to run Jais 3 on a Mac

There are two first-party paths. Pick by the app you already use.

llama.cpp or LM Studio (GGUF). Support for the K2 Horizon architecture was merged into mainline llama.cpp on 6 October 2026 (PR 29535), so any recent build should load IFM’s GGUF files. The cards ask for --jinja so the chat template is applied.
MLX. IFM publishes 4-bit, 8-bit and BF16 MLX repos for the 0.9B to 36B models. These are the native Apple silicon path.
Ollama. There is no Jais page in the Ollama library: the jais, jais3 and jais-3 paths all returned 404 on 4 November. The GGUF cards say the files work with llama.cpp and Ollama, which would mean pulling the Hugging Face file directly. I have not confirmed that.

llama.cpp: the 7B Q4_K_M GGUF straight from IFM’s repo, with the chat template

llama-server -hf IFM/Jais-3-7B-GGUF:Q4_K_M --jinja

The model cards suggest these sampling settings. Temperature 0.2 by default; 1.0 for English with thinking on, and 0.7 for Arabic with thinking on. Top_p 0.95 and top_k 20. Reasoning effort defaults to “high”, which costs tokens, so lower it for short lookups.

If you are choosing between apps, my LM Studio vs Ollama note covers how each one handles templates and structured output on a Mac.

What the published scores say

All of these numbers come from IFM. I found no independent Arabic evaluation of Jais 3 as of 4 November.

Vendor scores from the Jais 3 model cards
Comparison	Test	Jais 3	Rival
Jais-3-7B vs Qwen3.5-9B	Arabic MMLU	85.2	80.9
Jais-3-32B vs Qwen3.8-27B	Arabic MMLU	88.6	85.8
Jais-3-32B vs Qwen3.8-27B	Terminal Bench 2.1	63.7	79.8

Jais leads on Arabic knowledge and trails Qwen on coding agent work. IFM’s blog also says the 7B averages 79.6 on its Arabic evaluations against 56.4 for the K2 Horizon base. If your work is English code, Qwen 3.8 27B is still the stronger choice in this size class. If it is Arabic documents, Jais is the first open family I would test.

Jais 3 or Falcon for Arabic?

TII’s Falcon-Emirati is the other UAE Arabic model people ask about. On 4 November I found no Falcon-Emirati weights on Hugging Face, so I could not compare file sizes or local support. Jais 3 is the one you can download and run today.

FAQ: running Jais 3 locally

Can I run Jais 3 on a 16 GB Mac?

Yes, up to the 7B. Its Q4_K_M GGUF is 5.59 GB and the MLX 4-bit build is 7.2 GB, both comfortable on 16 GB. The 32B and MoVA models need 48 GB in my planning.

Is Jais 3 on Ollama?

Not in the official library as of 4 November 2026. IFM’s GGUF files run in llama.cpp, and the cards say they also work with Ollama, which I have not checked.

What license is Jais 3 under?

Apache 2.0, according to IFM’s blog and the license field on each Hugging Face repo.

Does Jais 3 speak Emirati Arabic?

IFM says it was trained on Modern Standard Arabic, dialects including Emirati and Egyptian, and Arabizi. I have not tested dialect quality.

Which Jais 3 should I pick for a 48 GB Mac?

Start with MoVA-36B-A4B at Q4_K_M (22.4 GB). It uses about 4B parameters per token, so it should be faster than the dense 32B at the same memory. If speed is fine with the 32B, its dense weights are the safer bet for quality.

Related local notes

Sources and what I did not test

I have not downloaded or run Jais 3. Every size here is a vendor file size read on 4 November 2026, and every score is IFM’s.

Sources

Suhail Mohebi · published 4 November 2026 · the cover is a generated illustration