We're building open-weight foundation models and the systems that refine and improve them.
Laguna Frontier-class reasoning at mid-size cost. Optimized for quality, speed, and efficiency. 118B params 8B active 1M context
Laguna Our lightest and fastest agentic coding model. Small enough to run on-device. 33B params 3B active 256K context
We work in the open. A closer look at the research, the engineering, and the ideas behind our models.
The Model Factory series A six-part blog series on how we train, evaluate, and iterate on foundation models. From GPU-to-GPU weight transfers and automated architecture ablations to reinforcement learning from code execution at scale.
Through the looking glass of benchmark hacking Outlining some of the reward hacks we’ve encountered and what strategies we are exploring to resolve them.
Laguna XS.2 and M.1: A Deeper Dive We've released the first two models in the Laguna family, Laguna M.1 and Laguna XS.2, alongside the runtime we use to train and operate agents, available through two product experiences in research preview.
Tools of the Trade: C2C Activation Offloading on Grace Blackwell We demonstrate the potential of NVIDIA's NVLink C2C on Grace-based superchips as a high-performance alternative to selective activation checkpointing. By offloading MLP activations to host memory during training, we achieve a 6–13% throughput improvement over selective AC with negligible memory overhead.
Go further with us. From real-world missions to the path to AGI, and the people building it.