Introducing Mercury 2.5
Inception Labs, 2026
Inception's flagship diffusion LLM, delivering cost-optimized frontier intelligence at 1107 tok/s. I contributed to the post-training, evaluation, and release of the model.
Adarsh Iyer
I study Computer Science at MIT and do research on vision and generative modeling in Professor Vincent Sitzmann's group. I have trained and shipped state-of-the-art LLMs through internships at Inception Labs and Continue (acquired by Cursor).
I'm broadly interested in robot learning, generative modeling, reinforcement learning, perception, and open-ended exploration. I also enjoy running cross country and track for MIT, playing music, painting, woodworking, and discussing philosophy.
News
- Mercury 2.5 is released, the leading diffusion LLM!
- We released Instinct, the state-of-the-art open model for next-edit prediction!
Work
Lights Are Secretly Cameras: Simulating Dual Photography in Mitsuba 3
MIT 6.4400 (Computer Graphics), 2025 · Top 5 project out of 44
A simulation of Sen et al., using a light as a camera, and a camera as a light. The adaptive multiplexing algorithm speeds up acquisition of the light transport matrix by 1000x.
Introducing Instinct: the world's best open Next Edit model, built by Continue
Continue, 2025
A next-edit prediction model for coding. I got to execute the full development cycle, from securing compute resources and running experiments to evaluations and release.
Expressive Symbolic Regression for Interpretable Models of Discrete-Time Dynamical Systems
arXiv, 2024
A custom neural network with a new fine-tuning scheme to identify simple mathematical expressions for chaotic discrete-time dynamical systems.
gAIt: Deep Learning Based Evaluation of Injurious Running Biomechanics Using Pose Estimation
IEEE ICPRS, 2023 · Runner-Up Student Paper
A high school research project on classifying running biomechanics from pose time series.