Qehwa AI: Pakistani Developer Creates World’s First Pashto AI LLM and Chatbot

A new large language model, Qehwa, has been developed by Junaid Ahmed, in a solo effort, to serve more than 60 million Pashto speakers worldwide. Inspired by Qalb, the Urdu LLM built by Taimoor Hassan, the model is designed specifically for the Pashto’s Peshawari dialect and aims to address limitations seen in existing global AI systems when handling the Pashto language and cultural context.

The model was evaluated using a custom benchmark consisting of 150 tests across 15 categories, where it achieved an overall accuracy of 85.3%. This makes it the first AI model built specifically for Pashto with structured evaluation benchmarks.

Training Process

Qehwa was developed in two main stages to improve both language understanding and task performance.

Qwen2.5-7B is the foundation of Qehwa. Qwen is a family of highly capable, open-source large language models developed by Alibaba Cloud. The “7B” means it has 7 billion parameters (the “brain cells” of the model). The developer took this base model—which already had a deep understanding of general logic, coding, and several languages—and specifically trained it to master the Peshawari Pashto dialect.

In the first stage, the model underwent continued pre-training using 3.4 million Pakistani Pashto documents. This phase focused on improving vocabulary, grammar, and cultural understanding.

In the training details, it mentions a “LoRA rank” of 64. LoRA is a clever mathematical technique that allows developers to fine-tune massive AI models without needing a room full of supercomputers. Instead of updating all 7 billion parameters, LoRA only updates a small, crucial subset, making the training process affordable and efficient for a solo developer.

In the second stage, the model was fine-tuned on more than 100,000 Pashto instruction pairs. This allowed it to follow prompts, answer questions, perform translations, and handle conversational tasks.

Capabilities and Benchmark Performance

The model supports prompts in Pashto, English, and Urdu, while generating responses in pure Pashto. It also introduces the first dedicated Pashto LLM benchmark with 150 evaluation tests.

Performance across key benchmarks includes:

English to Pashto translation reached 90% accuracy, while Urdu to Pashto translation scored 84%. In subject-specific categories, the model achieved 90% in culture and history, health and daily life, and geography and nature. The overall score across all 15 categories stands at 85.3%.

Independent Development and Open Source Access

The project was built independently without funding, a team, or institutional backing. The entire process, including dataset creation, training pipelines, debugging Unicode-related issues, managing GPU failures, and running multiple training cycles, was handled by a single developer.

The developer acknowledged guidance from Faiza Ghaffar during the project. Support from the open-source ecosystem, including Unsloth AI and Hugging Face, also played a role in enabling development.

The model is available as a free and open-source project, allowing researchers and developers to explore and build upon it.

Different Ways to Install and Run the Model

Unsloth: A popular open-source tool that makes fine-tuning and running LLMs significantly faster (up to 2x faster inference) and less memory-heavy. You can find more details on installation and running here.

BitsAndBytes (4-bit Quantization): This is a compression technique. A 7B model normally requires a lot of expensive video memory (VRAM). Quantization shrinks the model down so people can run it on regular consumer graphics cards (like an 8GB gaming GPU) instead of needing enterprise-grade server hardware.