Instructions to use CULUREYES-stack/Operator-4B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use CULUREYES-stack/Operator-4B with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="CULUREYES-stack/Operator-4B", trust_remote_code=True) messages = [ {"role": "user", "content": "Who are you?"}, ] pipe(messages)# pip install -U transformers accelerate # Load model directly from transformers import AutoModelForCausalLM model = AutoModelForCausalLM.from_pretrained("CULUREYES-stack/Operator-4B", trust_remote_code=True, device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use CULUREYES-stack/Operator-4B with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "CULUREYES-stack/Operator-4B" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "CULUREYES-stack/Operator-4B", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/CULUREYES-stack/Operator-4B
- SGLang
How to use CULUREYES-stack/Operator-4B with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "CULUREYES-stack/Operator-4B" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "CULUREYES-stack/Operator-4B", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "CULUREYES-stack/Operator-4B" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "CULUREYES-stack/Operator-4B", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Docker Model Runner
How to use CULUREYES-stack/Operator-4B with Docker Model Runner:
docker model run hf.co/CULUREYES-stack/Operator-4B
Download README.md from CULUREYES-stack/Operator-4B: direct link, hf CLI and curl.
- Browser
- Download file 7.3 kB
-
https://huggingface.co/CULUREYES-stack/Operator-4B/resolve/refs%2Fpr%2F2/README.md
- Command line
-
hf download hf://CULUREYES-stack/Operator-4B@refs/pr/2/README.md
-
curl -L -H "Authorization: Bearer $HF_TOKEN" -o README.md https://huggingface.co/CULUREYES-stack/Operator-4B/resolve/refs%2Fpr%2F2/README.md
language:
- en
- ar
pipeline_tag: text-generation
tags:
- operator
- hajerilm
- cultureyes
- arabic
- english
Operator 4B
Operator is an Arabic-English model from the HajeriLM family, published by Cultureyes.
العربية
Operator 4B نموذج لغوي عربي وإنجليزي من عائلة HajeriLM، تنشره Cultureyes. أشارك أوزانه ليتمكن الباحثون والمطورون من فحصها، ومواصلة التدريب، وتخصيص النموذج، والبناء عليه. طُوّر هذا العمل المستقل في الإمارات مع التركيز على الاستفادة الجيدة من وقت الحوسبة والموارد المتاحة.
I’m sharing this work in the spirit of the UAE’s commitment to innovation. I built it by working with the compute I had and repeatedly asking how to make each training step count. I’m sharing the weights so other people can inspect them, continue training, specialize them, and build their own work on top.
The model is an open starting point from an independent effort in the UAE, shared for people who want to explore what they can build from it.
A personal note
I’m sharing Operator during UAE Innovates 2026, as my own contribution from the UAE. This is independent work, not a government project or a claim that one model can match the resources of a major lab. I built with the compute I had, and I’m releasing the weights so others can inspect them, continue training, specialize them, and build something of their own.
I keep coming back to President Sheikh Mohamed bin Zayed’s description of an Emirati who talks about AI and still gathers firewood and makes coffee in the desert: “This is the true Emirati.” At the Union Parade, an Emirati participant recalled the President’s words: “all tribes in the UAE are one tribe.” Sheikh Khaled bin Mohamed bin Zayed put the scale of our ambition simply: “But we punch above our weight.” That is the spirit I hope this small independent effort carries.
ملاحظة شخصية
أشارك Operator خلال الإمارات تبتكر 2026 بوصفه مساهمة شخصية من الإمارات. هذا عمل مستقل، وليس مشروعًا حكوميًا أو ادعاءً بأن نموذجًا واحدًا يوازي إمكانات المختبرات الكبرى. بنيته بالحوسبة المتاحة لي، وأنشر أوزانه كي يتمكن الآخرون من فحصها ومواصلة تدريبها وتخصيصها والبناء عليها.
أعود كثيرًا إلى وصف صاحب السمو الشيخ محمد بن زايد للمواطن الإماراتي الذي يتحدث عن الذكاء الاصطناعي، ويحافظ في الوقت نفسه على تقاليد البر والقهوة: «هذا هو الإماراتي الحقيقي». وفي مسيرة الاتحاد، نقل أحد المشاركين عن سموه قوله إن «كل القبائل في الإمارات قبيلة واحدة». وقال سمو الشيخ خالد بن محمد بن زايد عن طموح الدولة: «لكننا نحقق أثرًا يتجاوز حجمنا» (ترجمة للعبارة الإنجليزية). هذه هي الروح التي أتمنى أن يحملها هذا العمل المستقل الصغير.
Languages
Operator was trained for use in English and Arabic. The model card and examples should be read with that scope in mind; this statement does not imply equal performance across languages or tasks.
دُرّب Operator للاستخدام باللغتين العربية والإنجليزية. ولا يعني ذلك تكافؤ الأداء في اللغتين أو في جميع المهام.
Model
- Model family: HajeriLM
- Publisher: Cultureyes
- Format: BF16 safetensors
- Parameter scale: approximately 4B
- Architecture: 32 layers, hidden size 2048, 16 attention heads, 2 key/value heads
- Position encoding: RoPE with theta 500,000; maximum configured context 32,768 tokens
- Attention gates: enabled
- License: not specified; reuse terms have not been declared
يتكوّن النموذج من 32 طبقة، وبُعد مخفي قدره 2048، ويستخدم 16 رأس انتباه ورأسي مفتاح/قيمة. يعتمد RoPE بقيمة theta تساوي 500,000، مع سياق أقصى مضبوط على 32,768 رمزًا، وبوابات انتباه مفعّلة.
Local inference
Two GGUF files are available in this repository:
- BF16 GGUF — full-precision bfloat16 weights.
- Q8_0 GGUF — 8-bit quantized weights for a smaller download and memory footprint.
These files use the Hajeri model architecture. They require a llama.cpp build that includes Hajeri architecture support; stock llama.cpp builds may not recognize them. The conversion was checked against the source checkpoint and tokenizer. The Q8_0 file is quantized, so its values are not bit-for-bit identical to BF16.
يوجد في هذا المستودع ملفان بصيغة GGUF:
تستخدم هذه الملفات معمارية Hajeri، وتتطلب إصدارًا من llama.cpp يدعم هذه المعمارية؛ قد لا تتعرف عليها الإصدارات القياسية. تمت مراجعة التحويل مقارنةً بملف النموذج الأصلي ومقسّم الرموز. نسخة Q8_0 مكمّمة، لذلك لا تطابق قيم BF16 عنصرًا بعنصر.
Training approach
The training workflow was built around a practical question: how can limited compute be used more carefully? The approach tracks cross-entropy per example, can stop spending updates on examples below a chosen loss threshold, saves resumable checkpoints, and adapts the learning rate from training loss. The aim is to focus training effort where it is still useful and make model building more accessible beyond large clusters.
Intended use
Use this checkpoint for research, experimentation, continued pretraining or fine-tuning, and local model development. Test it on your own tasks before relying on it in a product or consequential setting.
The weights are publicly downloadable, but no license grant is declared in this release. Reuse and redistribution terms remain unresolved.
Citation and attribution
Please refer to this model as Operator, from the HajeriLM family, published by Cultureyes. Preserve the model name and source link when sharing derivatives.
Release notes
Further development notes and the journal will document how the work evolves.