Zuman.ai

Zuman.ai is a multilingual conversational assistant designed to understand users and respond naturally in their language. It focuses on Ethiopian and African language use, with support intended for:

• Amharic (am)
• Afaan Oromo (om)
• Tigrinya (ti)
• Somali (so)
• English (en)

The model is based on a merged Gemma 4 BF16 fine-tune. It is intended for multilingual conversation and culturally aware responses. It may also be used for CBE-oriented assistance, but answers about banking products, policies, or procedures should be checked against current official sources.

Model details

Property	Details
Model name	Zuman.ai
Architecture	Gemma4UnifiedForConditionalGeneration
Checkpoint format	Safetensors, sharded
Weight dtype	BF16
Fine-tuning	Merged fine-tune; source folder identifies LoRA rank 64
Intended languages	Amharic, Afaan Oromo, Tigrinya, Somali, English

The configuration also includes multimodal components. The examples below demonstrate text-only chat; image, audio, and video behavior has not been independently evaluated here.

Quick start

This checkpoint is approximately 23 GB on disk. Loading it in BF16 requires a BF16-capable accelerator and additional memory for runtime state. A GPU with roughly 32 GB or more of memory is a practical starting point; actual requirements depend on the Transformers version, input length, and generation settings.

Install PyTorch for your platform and a Transformers release that supports Gemma4UnifiedForConditionalGeneration (the checkpoint configuration was written with Transformers 5.17.0):

pip install "transformers>=5.17.0" accelerate

Then load the model and send a chat:

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "DawitShibabaw/zuman.ai"

processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
model.eval()

messages = [
    {
        "role": "system",
        "content": (
            "You are Zuman.ai, a multilingual AI assistant designed for "
            "Ethiopian and African languages. Understand the user's language "
            "and respond naturally in the same language. Support Amharic, "
            "Afaan Oromo, Tigrinya, Somali, and English. Respect the user's "
            "cultural and linguistic context."
        ),
    },
    {
        "role": "user",
        "content": "ሰላም፣ ኢትዮጵያ ስለምን ታዋቂ ናት?",
    },
]

prompt = processor.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = processor(text=[prompt], return_tensors="pt").to(model.device)

with torch.inference_mode():
    output_ids = model.generate(
        **inputs,
        max_new_tokens=256,
        do_sample=False,
    )

new_tokens = output_ids[:, inputs["input_ids"].shape[-1] :]
print(processor.batch_decode(new_tokens, skip_special_tokens=True)[0])

The wording of any generated answer can vary. The prompt above asks the model to answer in Amharic because the user wrote in Amharic.

Serving with vLLM

The model can also be served through vLLM when using a release that supports the Gemma4UnifiedForConditionalGeneration architecture and this checkpoint's chat template. For example:

vllm serve DawitShibabaw/zuman.ai \
  --dtype bfloat16 \
  --max-model-len 8192 \
  --trust-remote-code \
  --chat-template chat_template.jinja

If the chat template is not available in the current directory, download the repository files first or provide the path to the downloaded chat_template.jinja. Check the vLLM documentation for compatible versions and hardware requirements.

Prompting

Provide a system message that states the assistant's role, desired language behavior, and any domain-specific boundaries. For example:

You are Zuman.ai, a multilingual AI assistant designed for Ethiopian and
African languages. Understand the user's language and respond naturally in
the same language. Support Amharic, Afaan Oromo, Tigrinya, Somali, and English.
Respect the user's cultural and linguistic context.

Use the bundled tokenizer/processor and chat template rather than manually constructing Gemma control tokens.

Intended use

• Multilingual conversation and question answering in the listed languages.
• Language-aware assistance for Ethiopian and African contexts.
• Prototyping and research into multilingual assistant applications.
• CBE-oriented assistance when used with appropriate, current, authoritative source material.

Limitations

• Language quality may vary by language, dialect, spelling, and subject.
• The model may produce inaccurate, outdated, or culturally inappropriate responses. Verify important claims with reliable sources.
• Do not use generated output as a substitute for professional financial, legal, medical, or other high-stakes advice.
• This model card does not provide benchmark results or a validated comparison of performance across the supported languages.
• Training-data composition, evaluation methodology, and quantitative evaluation results are not documented here.
• This repository does not specify a license. Review the applicable Gemma 4 terms and confirm redistribution and usage rights before using or sharing the checkpoint.

Citation

If you use or share this model, please cite the model repository:

DawitShibabaw. Zuman.ai. Hugging Face model repository.
https://huggingface.co/DawitShibabaw/zuman.ai