Rameswar Panda

Research

Most of what I work on tries to answer a simple question: how much capability can you get out of a given amount of compute? In practice that means building open models, finding architectures that are cheaper to train and serve, and working out training recipes and data that hold up at scale.

Granite: Granite is IBM's family of open language models, released under Apache 2.0. Since Granite Code in 2024, the team has shipped Granite 3.0 through 4.2, adding long context, switchable reasoning, a hybrid Mamba-2/Transformer architecture, Nano models as small as 350M, and built-in tool calling. Granite received the highest score in the history of Stanford's Foundation Model Transparency Index (95/100 in 2025).

Efficient architectures: With collaborators at MIT, we develop alternatives to standard softmax attention, including Gated Linear Attention, PaTH attention and stick-breaking attention, along with ways to shrink the KV cache and distill Transformers into hybrid models.

Training recipes: Hyperparameters and data that work at one scale often break at another. Our work here includes the Power Scheduler, which needs no retuning as batch size or token count changes, CodeAlchemy for synthetic code data, and PRISM, which shows that a short mid-training phase is what makes later reinforcement learning work.

Agents: We built the agentic RL infrastructure and environments behind Granite 4.2, training models in real software repositories, live terminals, and web search. Tool use is also limited by data: Toucan, our open dataset of 1.5M trajectories from real MCP environments, is used to train many other open models.

Vision (earlier work): Before language models, I worked on efficient computer vision, including CrossViT, adaptive video recognition and multi-task learning.

Open source

Most of our work is released openly. Some of the projects people use the most:

Publications

The full list of my publications is on Google Scholar. Here is a short selection:

Press

Service