ALLaM-7B-Instruct-preview

ALLaM-7B-Instruct-preview

Answers questions and follows instructions in Arabic and English. A 7B model trained from scratch by the National Center for Artificial Intelligence at SDAIA, for developers and researchers building Arabic language tools.

Made by humain-ai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: ALLaM-7B-Instruct-preview is a 7B instruction-tuned language model for Arabic and English, built by the National Center for Artificial Intelligence at SDAIA. It was trained from scratch, first on 4T English tokens and then on 1.2T mixed Arabic and English tokens, a plan the card says keeps its English skills intact. It aims to speed up research and product work in Arabic language technology.

Who it suits: Developers and researchers who build Arabic or bilingual assistants and want a base to test and extend. The weights carry the Apache 2.0 license.

Industry | Who uses it | How they use it | Example
Language technology | Researchers | Study and extend Arabic model capabilities | Testing question answering in Arabic
Customer service | Product teams | Build bilingual Arabic and English assistants | A support bot that answers in Arabic
Content | Content teams | Draft Arabic explanations and answers | Explaining a household task in Arabic

Anyone who wants a finished, safety-checked product can skip it, because the card says the model is one part of a larger system and developers must add their own safety measures.

Where to use it:

• Python with the transformers library, which the card tested at version 4.40.1 or newer
• A CUDA GPU, which the card's example loads the model onto
• Windows, macOS or Linux with Python, for downloading the files

Use case | Who | What they get back
Arabic Q and A | App builders | Answers written in Arabic
Bilingual chat | Product teams | Replies in Arabic or English
Research baseline | Researchers | A from-scratch Arabic and English model to test

How to set it up:

• Python and the transformers package, version 4.40.1 or newer
• A CUDA GPU
• Room for the weights, about 14 GB in BF16

1. Install Python and a transformers version of 4.40.1 or newer.
2. Install the huggingface_hub package with the first command below.
3. Download the model files with the second command.
4. Pick the older release or the newest one through the revision setting, v1 or v2, if you need a specific one.

pip install -U huggingface_hub
hf download humain-ai/ALLaM-7B-Instruct-preview

How to use it: The card's snippet loads the model and tokenizer, wraps an Arabic question in the chat template and generates a reply on the GPU. It names the model with the id ALLaM-AI/ALLaM-7B-Instruct-preview, so swap in the repository you downloaded if the two differ.

from transformers import AutoModelForCausalLM, AutoTokenizer
allam_model = AutoModelForCausalLM.from_pretrained("ALLaM-AI/ALLaM-7B-Instruct-preview")
tokenizer = AutoTokenizer.from_pretrained("ALLaM-AI/ALLaM-7B-Instruct-preview")
messages=[
    {"role": "user", "content": "كيف أجهز كوب شاهي؟"},
]
inputs = tokenizer.apply_chat_template(messages, tokenize=False)
inputs = tokenizer(inputs, return_tensors='pt', return_token_type_ids=False)
inputs = {k: v.to('cuda') for k,v in inputs.items()}
allam_model = allam_model.to('cuda')
response = allam_model.generate(**inputs, max_new_tokens=4096, do_sample=True, top_k=50, top_p=0.95, temperature=.6)
print(tokenizer.batch_decode(response, skip_special_tokens=True)[0])

The model works best with no preset system prompt, though you can add your own, in English or Arabic. The card gives a short line that tells it to act as a bilingual English and Arabic assistant. The system prompt lives in the tokenizer config and is reached through the chat template.

Example: You ask in Arabic how to prepare a cup of tea, as the card's snippet does. The model answers in Arabic, with up to 4,096 new tokens allowed.

For newcomers and experts: This is a base for building Arabic tools, not a finished chat app. A good first step is to run the snippet above with the Arabic question and read the reply.

The card describes two training paths for the ALLaM series, pretraining from scratch and continued training from open-weight models, and this release follows the first. It was trained on 5.2 trillion tokens, with a codebase built on NVIDIA Megatron-LM, an average MFU of about 42 percent and bf16-mixed precision. Post-training used 7M instructions and 260K preference pairs, and the context length is 4096 tokens.

Pretrain on 4T English tokens
Pretrain on 1.2T Arabic and English tokens
Tune on 7M instructions
Tune on 260K preference pairs
Chat in Arabic or English

Good to know:

• License: Apache 2.0 in Hugging Face's record, and the card points to its LICENSE file
• Cost to run: none named
• Data it reaches: the download fetches files from Hugging Face, and the source names nothing else
• It is a preview release, and the card asks developers to add safety measures, especially when connecting it to external tools

Title: Pretraining tokens by stage, in trillions
English stage: 4 T
Arabic and English stage: 1.2 T

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 7B
• Weights: about 14 GB in BF16; about 3.5 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 93,626
• Likes: 207
• License: apache-2.0
• Languages: ar, en
• Quantized versions: 15
• Fine-tunes: 29
• Tags: transformers, safetensors, llama, text-generation, pytorch, conversational

Links

• Model: https://huggingface.co/humain-ai/ALLaM-7B-Instruct-preview
• Files: https://huggingface.co/humain-ai/ALLaM-7B-Instruct-preview/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download humain-ai/ALLaM-7B-Instruct-preview

Then follow the usage in its model card. Nothing in this kit runs by itself.