bf16 MLX conversion of IFM/Jais-3-32B, converted from revision 847c6fc. 69.6 GB on disk. For Apple silicon.
Jais-3-32B is the large dense Jais-3 model: a 32B decoder-only model with a 512K (524,288-token) context window.
mlx-lm doesn't support the k2_horizon architecture (used by Jais-3) yet. Until support lands, this repo ships the MLX model code (k2_horizon.py), which mlx-lm loads through the model_file entry in config.json. So the released mlx-lm works as-is:
pip install -U mlx-lm
Pass --trust-remote-code (or trust_remote_code=True). The file runs on your machine, so read it first if you like; its comments describe how it differs from the PyTorch code.
No quantization: the original bf16 weights in MLX format.
Checked against the PyTorch code (modeling_k2_horizon.py) in fp32 on the real weights of this model. The check used the model truncated to its first 4 layers (plus the final norm and lm_head), with the same truncation on both sides, on English, code, Arabic and math prompts: the logits match to a relative error of 2e-6 or better, the next-token predictions agree at every position, and greedy decoding with the KV cache gives the same tokens (64/64 on every prompt). All 64 layers run the same code path as the smaller dense Jais-3 models, which were checked in full.
Smoke-tested after conversion with released mlx-lm 0.32.0 on English and Arabic chat prompts: answers end normally (the chat template and stop tokens work).
Perplexity on 128 × 512 tokens (WikiText-2 test, and the first articles of the Arabic Wikipedia dump 20231101.ar; lower is better) and generation speed on an Apple M4 Max, all measured the same way:
| bf16 | 8-bit | 4-bit | |
|---|---|---|---|
| Bits/weight | 16.0 | 8.5 | 5.9 |
| Disk | 69.6 GB | 37.0 GB | 25.5 GB |
| Peak memory | n/a | 37.2 GB | 25.8 GB |
| WikiText-2 perplexity (English) | n/a | 9.324 | 9.521 (+2.1%) |
| Arabic Wikipedia perplexity | n/a | 11.307 | 11.661 (+3.1%) |
| Generation | n/a | 12.4 tok/s | 18.8 tok/s |
The bf16 configuration could not be tested; the percentages are relative to 8-bit, which is within 0.2% of bf16 on every smaller Jais-3 model.
Perplexity is a coarse signal. Test the versions on your own workload before picking one.
mlx_lm.generate --model IFM/Jais-3-32B-MLX-bf16 --trust-remote-code --prompt "ما هي عاصمة الإمارات؟" --max-tokens 2048
from mlx_lm import load, generate
model, tokenizer = load("IFM/Jais-3-32B-MLX-bf16", trust_remote_code=True)
messages = [{"role": "user", "content": "Explain mixture-of-experts in two sentences."}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
print(generate(model, tokenizer, prompt, max_tokens=2048))
mlx_lm.chat and mlx_lm.server take the same --trust-remote-code flag.
The model thinks before it answers, inside <ifm|think> … </ifm|think>. Leave room for that in max_tokens. The effort level is set with reasoning_effort in the chat template: "high" (default), "medium" or "low", e.g. apply_chat_template(messages, add_generation_prompt=True, reasoning_effort="low").
Notes:
<ifm|think> tags yet, so mlx_lm.server returns the thinking text inside content rather than in a separate reasoning field. The tool-call format isn't parsed yet either.Apache-2.0, inherited from the base model. Refer to the original model card for architecture, benchmarks and intended use.
@misc{ifm2026jais3introducingarabiccentric,
title={Jais 3: Introducing Arabic-centric Agentic AI},
author={Team IFM and Haonan Li and Hadi Abdine and Nikhil Ranjan and Mohamed Anwar and Ahmad Chamma and Amr Mohamed and Ahmed Frikha and Dani Bouch and Maaz Maqsood and George Ibrahim and Mostafa Awad and Nurkhan Laiyk and Hasan Iqbal and Abhishek Maiti and Bilal Bouardi and Shivam Chauhan and Ahmer Tabassum and Owais Aijaz and Omar Arbi and Neha Sengupta and Aaryamonvikram Singh and Abdelhakim Freihat and Sarath Chandran and Rahul Pal and Mohammed Bellam and Sajid Siddiki and Sarah AlBarri and Debopriyo Banerjee and Ali El Filali and Samta Kamboj and Rifo Genadi and Ahmed Attia and Go Inoue and Shuonan Dong and Amine Khelif Khelif and Evan Dufraisse and Florian Valade and Tapas Sen and Guowei He and Sunil Sahu and Onkar Pandit and Kareem Magdy and Sheikh Khalid bin Khalifa Al Nahyan and Samujjwal Ghosh and Zainul Abedien and Ahmed Quraishi and Larry Murray and Avraham Sheinin and Natalia Vassilieva and Preslav Nakov and Guokan Shang and Xuguang Hector Ren and Zhengzhong Liu and Eric Xing},
year={2026},
eprint={2611.10224},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2611.10224},
}
model-00001-of-00014.safetensors 5.3 GBmodel-00006-of-00014.safetensors 5.3 GBmodel-00010-of-00014.safetensors 5.3 GBmodel-00002-of-00014.safetensors 5.3 GBmodel-00009-of-00014.safetensors 5.3 GBmodel-00008-of-00014.safetensors 5.3 GBmodel-00004-of-00014.safetensors 5.3 GBmodel-00012-of-00014.safetensors 5.3 GBmodel-00005-of-00014.safetensors 5.3 GBmodel-00011-of-00014.safetensors 5.2 GBmodel-00007-of-00014.safetensors 5.2 GBmodel-00003-of-00014.safetensors 5.2 GBmodel-00013-of-00014.safetensors 3.6 GBmodel-00014-of-00014.safetensors 2.6 GB有文件?提交种子。
提交种子inception42/Jais-3-32B