AraFusion โ Arabic Masked Diffusion Language Model
Trained with MDLM on FineWeb-2 Arabic subsets. Supports dialect-conditioned generation via Classifier-Free Guidance (CFG).
Model details
| Architecture | DiT (BERT-base scale: 12L / 12H / 768d) |
| Vocab | 96 000 (MorphBPE, see AraFusion/arafusion-morphBPE) |
| Sequence length | 512 tokens |
| Diffusion | Linear noise schedule, 1 000 sampling steps |
| Dialects | MSA ([MSA]), Najdi ([NAJDI]), Egyptian ([EGYPT]) |
| CFG p_uncond | 0.10 |
| Training steps | 200 000 (pre-train) |
| Precision | BF16 |
Data
Trained on three FineWeb-2 Arabic subsets:
| Subset | Dialect | Split |
|---|---|---|
arb_Arab |
Modern Standard Arabic | train 10 % |
ars_Arab |
Najdi / Saudi Arabic | train (full) |
arz_Arab |
Egyptian Arabic | train (full) |
See AraFusion/arafusion-arabic-raw for the raw text and
AraFusion/arafusion-arabic-packed for the packed training sequences.
Checkpoints
Training job: vanillabpe-641m-51479
| Path in repo |
|---|
checkpoints/0-34500-v1.ckpt |
checkpoints/1-35000-v1.ckpt |
checkpoints/1-35500.ckpt |
checkpoints/1-36000.ckpt |
checkpoints/1-36500.ckpt |
checkpoints/1-37000.ckpt |
checkpoints/1-37500.ckpt |
checkpoints/1-38000.ckpt |
checkpoints/1-38500.ckpt |
checkpoints/1-39000.ckpt |
checkpoints/1-39500.ckpt |
checkpoints/1-40000.ckpt |
checkpoints/1-40500.ckpt |
checkpoints/1-41000.ckpt |
checkpoints/1-41500.ckpt |
checkpoints/1-42000.ckpt |
checkpoints/1-42500.ckpt |
checkpoints/1-43000.ckpt |
checkpoints/1-43500.ckpt |
checkpoints/1-44000.ckpt |
checkpoints/1-44500.ckpt |
checkpoints/1-45000.ckpt |
checkpoints/1-45500.ckpt |
checkpoints/1-46000.ckpt |
checkpoints/1-46500.ckpt |
checkpoints/1-47000.ckpt |
checkpoints/1-47500.ckpt |
checkpoints/1-48000.ckpt |
checkpoints/1-48500.ckpt |
checkpoints/1-49000.ckpt |
checkpoints/1-49500.ckpt |
checkpoints/1-50000.ckpt |
checkpoints/1-50500.ckpt |
checkpoints/1-51000.ckpt |
checkpoints/1-51500.ckpt |
checkpoints/1-52000.ckpt |
checkpoints/1-52500.ckpt |
checkpoints/1-53000.ckpt |
checkpoints/1-53500.ckpt |
checkpoints/1-54000.ckpt |
checkpoints/1-54500.ckpt |
checkpoints/1-55000.ckpt |
checkpoints/1-55500.ckpt |
checkpoints/1-56000.ckpt |
checkpoints/1-56500.ckpt |
checkpoints/1-57000.ckpt |
checkpoints/1-57500.ckpt |
checkpoints/1-58000.ckpt |
checkpoints/1-58500.ckpt |
checkpoints/1-59000.ckpt |
checkpoints/1-59500.ckpt |
checkpoints/1-60000.ckpt |
checkpoints/1-60500.ckpt |
checkpoints/1-61000.ckpt |
checkpoints/1-61500.ckpt |
checkpoints/1-62000.ckpt |
checkpoints/1-62500.ckpt |
checkpoints/1-63000.ckpt |
checkpoints/1-63500.ckpt |
checkpoints/1-64000.ckpt |
checkpoints/1-64500.ckpt |
checkpoints/1-65000.ckpt |
checkpoints/1-65500.ckpt |
checkpoints/1-66000.ckpt |
checkpoints/1-66500.ckpt |
checkpoints/1-67000.ckpt |
checkpoints/1-67500.ckpt |
checkpoints/1-68000.ckpt |
checkpoints/1-68500.ckpt |
checkpoints/1-69000.ckpt |
checkpoints/2-100000.ckpt |
checkpoints/2-100500.ckpt |
checkpoints/2-101000.ckpt |
checkpoints/2-101500.ckpt |
checkpoints/2-102000.ckpt |
checkpoints/2-102500.ckpt |
checkpoints/2-103000.ckpt |
checkpoints/2-103500.ckpt |
checkpoints/2-69500.ckpt |
checkpoints/2-70000.ckpt |
checkpoints/2-70500.ckpt |
checkpoints/2-71000.ckpt |
checkpoints/2-71500.ckpt |
checkpoints/2-72000.ckpt |
checkpoints/2-72500.ckpt |
checkpoints/2-73000.ckpt |
checkpoints/2-73500.ckpt |
checkpoints/2-74000.ckpt |
checkpoints/2-74500.ckpt |
checkpoints/2-75000.ckpt |
checkpoints/2-75500.ckpt |
checkpoints/2-76000.ckpt |
checkpoints/2-76500.ckpt |
checkpoints/2-77000.ckpt |
checkpoints/2-77500.ckpt |
checkpoints/2-78000.ckpt |
checkpoints/2-78500.ckpt |
checkpoints/2-79000.ckpt |
checkpoints/2-79500.ckpt |
checkpoints/2-80000.ckpt |
checkpoints/2-80500.ckpt |
checkpoints/2-81000.ckpt |
checkpoints/2-81500.ckpt |
checkpoints/2-82000.ckpt |
checkpoints/2-82500.ckpt |
checkpoints/2-83000.ckpt |
checkpoints/2-83500.ckpt |
checkpoints/2-84000.ckpt |
checkpoints/2-84500.ckpt |
checkpoints/2-85000.ckpt |
checkpoints/2-85500.ckpt |
checkpoints/2-86000.ckpt |
checkpoints/2-86500.ckpt |
checkpoints/2-87000.ckpt |
checkpoints/2-87500.ckpt |
checkpoints/2-88000.ckpt |
checkpoints/2-88500.ckpt |
checkpoints/2-89000.ckpt |
checkpoints/2-89500.ckpt |
checkpoints/2-90000.ckpt |
checkpoints/2-90500.ckpt |
checkpoints/2-91000.ckpt |
checkpoints/2-91500.ckpt |
checkpoints/2-92000.ckpt |
checkpoints/2-92500.ckpt |
checkpoints/2-93000.ckpt |
checkpoints/2-93500.ckpt |
checkpoints/2-94000.ckpt |
checkpoints/2-94500.ckpt |
checkpoints/2-95000.ckpt |
checkpoints/2-95500.ckpt |
checkpoints/2-96000.ckpt |
checkpoints/2-96500.ckpt |
checkpoints/2-97000.ckpt |
checkpoints/2-97500.ckpt |
checkpoints/2-98000.ckpt |
checkpoints/2-98500.ckpt |
checkpoints/2-99000.ckpt |
checkpoints/2-99500.ckpt |
checkpoints/3-104000.ckpt |
checkpoints/3-104500.ckpt |
checkpoints/3-105000.ckpt |
checkpoints/3-105500.ckpt |
checkpoints/3-106000.ckpt |
checkpoints/3-106500.ckpt |
checkpoints/3-107000.ckpt |
checkpoints/3-107500.ckpt |
checkpoints/3-108000.ckpt |
checkpoints/3-108500.ckpt |
checkpoints/3-109000.ckpt |
checkpoints/3-109500.ckpt |
checkpoints/3-110000.ckpt |
checkpoints/3-110500.ckpt |
checkpoints/3-111000.ckpt |
checkpoints/3-111500.ckpt |
checkpoints/3-112000.ckpt |
checkpoints/3-112500.ckpt |
checkpoints/3-113000.ckpt |
checkpoints/3-113500.ckpt |
checkpoints/3-114000.ckpt |
checkpoints/3-114500.ckpt |
checkpoints/3-115000.ckpt |
checkpoints/3-115500.ckpt |
checkpoints/3-116000.ckpt |
checkpoints/3-116500.ckpt |
checkpoints/3-117000.ckpt |
checkpoints/3-117500.ckpt |
checkpoints/3-118000.ckpt |
checkpoints/3-118500.ckpt |
checkpoints/3-119000.ckpt |
checkpoints/3-119500.ckpt |
checkpoints/3-120000.ckpt |
checkpoints/3-120500.ckpt |
checkpoints/3-121000.ckpt |
checkpoints/3-121500.ckpt |
checkpoints/3-122000.ckpt |
checkpoints/3-122500.ckpt |
checkpoints/3-123000.ckpt |
checkpoints/3-123500.ckpt |
checkpoints/3-124000.ckpt |
checkpoints/3-124500.ckpt |
checkpoints/3-125000.ckpt |
checkpoints/3-125500.ckpt |
checkpoints/3-126000.ckpt |
checkpoints/3-126500.ckpt |
checkpoints/3-127000.ckpt |
checkpoints/3-127500.ckpt |
checkpoints/3-128000.ckpt |
checkpoints/3-128500.ckpt |
checkpoints/3-129000.ckpt |
checkpoints/3-129500.ckpt |
checkpoints/3-130000.ckpt |
checkpoints/3-130500.ckpt |
checkpoints/3-131000.ckpt |
checkpoints/3-131500.ckpt |
checkpoints/3-132000.ckpt |
checkpoints/3-132500.ckpt |
checkpoints/3-133000.ckpt |
checkpoints/3-133500.ckpt |
checkpoints/3-134000.ckpt |
checkpoints/3-134500.ckpt |
checkpoints/3-135000.ckpt |
checkpoints/3-135500.ckpt |
checkpoints/3-136000.ckpt |
checkpoints/3-136500.ckpt |
checkpoints/3-137000.ckpt |
checkpoints/3-137500.ckpt |
checkpoints/3-138000.ckpt |
checkpoints/3-138500.ckpt |
checkpoints/4-139000.ckpt |
checkpoints/4-139500.ckpt |
checkpoints/4-140000.ckpt |
checkpoints/4-140500.ckpt |
checkpoints/4-141000.ckpt |
checkpoints/4-141500.ckpt |
checkpoints/4-142000.ckpt |
checkpoints/4-142500.ckpt |
checkpoints/4-143000.ckpt |
checkpoints/4-143500.ckpt |
checkpoints/4-144000.ckpt |
checkpoints/4-144500.ckpt |
checkpoints/4-145000.ckpt |
checkpoints/4-145500.ckpt |
checkpoints/4-146000.ckpt |
checkpoints/4-146500.ckpt |
checkpoints/4-147000.ckpt |
checkpoints/4-147500.ckpt |
checkpoints/4-148000.ckpt |
checkpoints/4-148500.ckpt |
checkpoints/4-149000.ckpt |
checkpoints/4-149500.ckpt |
checkpoints/4-150000.ckpt |
checkpoints/4-150500.ckpt |
checkpoints/4-151000.ckpt |
checkpoints/4-151500.ckpt |
checkpoints/4-152000.ckpt |
checkpoints/4-152500.ckpt |
checkpoints/4-153000.ckpt |
checkpoints/4-153500.ckpt |
checkpoints/4-154000.ckpt |
checkpoints/4-154500.ckpt |
checkpoints/4-155000.ckpt |
checkpoints/4-155500.ckpt |
checkpoints/4-156000.ckpt |
checkpoints/4-156500.ckpt |
checkpoints/4-157000.ckpt |
checkpoints/4-157500.ckpt |
checkpoints/4-158000.ckpt |
checkpoints/4-158500.ckpt |
checkpoints/4-159000.ckpt |
checkpoints/4-159500.ckpt |
checkpoints/4-160000.ckpt |
checkpoints/4-160500.ckpt |
checkpoints/4-161000.ckpt |
checkpoints/4-161500.ckpt |
checkpoints/4-162000.ckpt |
checkpoints/4-162500.ckpt |
checkpoints/4-163000.ckpt |
checkpoints/4-163500.ckpt |
checkpoints/4-164000.ckpt |
checkpoints/4-164500.ckpt |
checkpoints/4-165000.ckpt |
checkpoints/4-165500.ckpt |
checkpoints/4-166000.ckpt |
checkpoints/4-166500.ckpt |
checkpoints/4-167000.ckpt |
checkpoints/4-167500.ckpt |
checkpoints/4-168000.ckpt |
checkpoints/4-168500.ckpt |
checkpoints/4-169000.ckpt |
checkpoints/4-169500.ckpt |
checkpoints/4-170000.ckpt |
checkpoints/4-170500.ckpt |
checkpoints/4-171000.ckpt |
checkpoints/4-171500.ckpt |
checkpoints/4-172000.ckpt |
checkpoints/4-172500.ckpt |
checkpoints/4-173000.ckpt |
checkpoints/5-173500.ckpt |
checkpoints/5-174000.ckpt |
checkpoints/5-174500.ckpt |
checkpoints/5-175000.ckpt |
checkpoints/5-175500.ckpt |
checkpoints/5-176000.ckpt |
checkpoints/5-176500.ckpt |
checkpoints/5-177000.ckpt |
checkpoints/5-177500.ckpt |
checkpoints/5-178000.ckpt |
checkpoints/5-178500.ckpt |
checkpoints/5-179000.ckpt |
checkpoints/5-179500.ckpt |
checkpoints/5-180000.ckpt |
checkpoints/5-180500.ckpt |
checkpoints/5-181000.ckpt |
checkpoints/5-181500.ckpt |
checkpoints/5-182000.ckpt |
checkpoints/5-182500.ckpt |
checkpoints/5-183000.ckpt |
checkpoints/5-183500.ckpt |
checkpoints/5-184000.ckpt |
checkpoints/5-184500.ckpt |
checkpoints/5-185000.ckpt |
checkpoints/5-185500.ckpt |
checkpoints/5-186000.ckpt |
checkpoints/5-186500.ckpt |
checkpoints/5-187000.ckpt |
checkpoints/5-187500.ckpt |
checkpoints/5-188000.ckpt |
checkpoints/5-188500.ckpt |
checkpoints/5-189000.ckpt |
checkpoints/5-189500.ckpt |
checkpoints/5-190000.ckpt |
checkpoints/5-190500.ckpt |
checkpoints/5-191000.ckpt |
checkpoints/5-191500.ckpt |
checkpoints/5-192000.ckpt |
checkpoints/5-192500.ckpt |
checkpoints/5-193000.ckpt |
checkpoints/5-193500.ckpt |
checkpoints/5-194000.ckpt |
checkpoints/5-194500.ckpt |
checkpoints/5-195000.ckpt |
checkpoints/5-195500.ckpt |
checkpoints/5-196000.ckpt |
checkpoints/5-196500.ckpt |
checkpoints/5-197000.ckpt |
checkpoints/5-197500.ckpt |
checkpoints/5-198000.ckpt |
checkpoints/5-198500.ckpt |
checkpoints/5-199000.ckpt |
checkpoints/5-199500.ckpt |
checkpoints/best.ckpt |
checkpoints/last.ckpt |
Usage
# Requires the MDLM library from https://github.com/kuleshov-group/mdlm
from transformers import PreTrainedTokenizerFast
tok = PreTrainedTokenizerFast.from_pretrained("AraFusion/arafusion-morphBPE")
# Load checkpoint and run a sampling step โ see MDLM docs for full API.
Citation
@misc{arafusion2026,
title = {AraFusion: Dialect-Conditioned Arabic Masked Diffusion Language Model},
year = {2026},
url = {https://huggingface.co/AraFusion}
}
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐ Ask for provider support