Twkeed-GPT-120B (توكيد)

A powerful Arabic language model based on GPT-OSS 120B (117B parameters, 5.1B active via MoE).

Fine-tuned on Saudi Arabian content including:

• Saudi Labor Law articles (Articles 53, 84-85, 109, 151)
• Saudi dialect understanding (يبي، وش، وين، الحين)
• Arabic grammar and writing
• Vision 2030 knowledge

Model Details

• Base Model: mlx-community/gpt-oss-120b-MXFP4-Q8
• Architecture: Mixture of Experts (117B total, 5.1B active)
• Reasoning: Near o4-mini level capabilities
• Fine-tuning Method: LoRA (r=8, alpha=16) with unsloth-mlx
• Training Hardware: Mac Studio M3 Ultra 96GB
• Language: Arabic (Modern Standard + Saudi Dialect)

Model Identity

This model identifies as توكيد (Twkeed) - an Arabic AI assistant with strong reasoning capabilities.

When asked "من أنت؟" (Who are you?), the model responds with its identity.

Usage

from mlx_lm import load, generate

model, tokenizer = load("twkeed-sa/twkeed-gpt-120b")

response = generate(
    model,
    tokenizer,
    prompt="مرحباً، من أنت؟",
    max_tokens=200,
)
print(response)

Why 120B?

Aspect	20B	120B
Parameters	21B (3.6B active)	117B (5.1B active)
Reasoning	Good	Excellent (near o4-mini)
Arabic Knowledge	Very Good	Excellent

The 120B base model provides much stronger reasoning and knowledge, requiring less fine-tuning adaptation.

Training Data

Fine-tuned on 40,000+ Arabic examples including:

• Arabic Alpaca dataset
• Custom Saudi Labor Law content
• Saudi dialect examples
• Arabic grammar instruction data

License

Apache 2.0

Author

Fine-tuned using unsloth-mlx