Ripple: A Local AI Model Co-Designed for GPUs

Today we're releasing Ripple, a frontier local model that outperforms US open-weight models fifteen X larger in size on agentic tasks like Terminal Bench at a size that can run locally on a MacBook Pro or DGX Spark.

By pre-training Ripple with a deep understanding of the hardware it runs on, we were able to reach the same performance as the open source baseline with sixteen X less compute.

It's like we trained with sixteen times more GPUs.

We built Ripple for three reasons. First, to show what we can gain by co-designing the model with the hardware it runs on, in this case, a GPU.

Second, to power our AI chip design platform.

An in-house version of Ripple orchestrates and learns across our proprietary AI algorithms to quickly design more performant chips.

Third, to start closing the loop, training better models that design better chips that train even better models.

Ripple is our first model, and it was trained on silicon that wasn't designed for it.

Imagine what's possible on silicon that is.