mann-e/Hormoz-8B

Hormoz 8B

Introduction

This model is an effort in order to make a multi-lingual and on device models which can be executed on the consumer hardware. The model follows the steps used in training DeepSeek model. However, the model is not a reasoning model and a generic question answering, conversational and uncensored model which has been made with a cost of around $4000 USD.

If you're curious about the model you also can see our GitHub and learn more about the benchmarks and costs.

Also, this model is based on Command R's architecture, since that architecture gave us the best results in multilingual chat. Specially with languages such as Persian and Arabic. This way, you can consider this model like a commercially useaeble version of aya expanse as well.

The name

The name Hormoz comes from the Persian word "هرمز" which has multiple meanings. It can point to the strait of Hormoz in Persian Gulf or Hormoz Island which is part of the Hormozgan Province in the south of Iran. Also it may point to "اورمزد" or Ourmozd which is middle/ancient Persian name for the planet Jupiter and derived from the term Ahura Mazda or the Avestan term for God.

How to run (transformers)

Install transformers

pip install transformers --upgrade

Note: For better performance, you may need to install accelerate package as well.

Inference

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "mann-e/Hormoz-8B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id).to("cuda")

messages = [{"role": "user", "content": "What is the answer to universe, life and everything?"}]
input_ids = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to("cuda")

gen_tokens = model.generate(
    input_ids, 
    max_new_tokens=1024, 
    do_sample=True, 
    temperature=1.0,
    )

gen_text = tokenizer.decode(gen_tokens[0])
print(gen_text)

License

This model is published under MIT license.

Commercial Use

Since this model is MIT licensed, you're free to do whatever you want with the model. However since we're a relatively small startup, we recommend you if you are a big corporate and you host this model, give us a capacity of your API as well. This way, we both can benefit from the model.