Why we started
DeltaTune Lab.
An independent AI research lab with a narrow promise: ask focused questions about efficient AI and publish what we learn.
DeltaTune Lab began with a simple observation: many of the most interesting questions in modern AI are not only about building bigger models. They are about getting more out of the models, data, and hardware we already have.
We are two Pakistan-based researchers, Rao Muhammad Dayan Atif and Syed Muhammad Sarim Ahsan, connected with Sir Syed University. We started the lab to give those questions a home, and to work on them carefully and in the open.
Three questions
Our research agenda is organised around three connected questions:
- How much really needs to change? Efficient fine-tuning and model adaptation.
- What makes a small model reason better? Post-training, reinforcement learning, and distillation for compact models.
- Where is the compute going? GPU memory movement, kernel fusion, and training bottlenecks.
They are connected on purpose. An adaptation method is only useful if it runs on accessible hardware, and a reasoning improvement matters more when we can explain what it cost.
What “independent” means to us
Being independent lets us choose our questions. It does not excuse us from rigor. We would rather publish one well-documented experiment than ten loosely supported claims. That is why our lab charter asks us to measure cost alongside capability, document our baselines, and report limits and failures.
What comes next
We are at the beginning. In the coming months this blog will carry explainers, experiment logs, and notes on releases as they become ready on Hugging Face. If you have a good question, or want to collaborate, write to us.