Willow taps CoreWeave Inc. for AI training support

The choice says as much about what Willow is not doing as what CoreWeave provides. Willow fine-tunes its own speech recognition model on quiet, real-world audio such as a user whispering in a coffee shop, and Liu described the post-processing step — a very small large language model that reformats what you say into what you would have written — as the main part of the company's 'secret sauce.' That is where the team's effort goes, which is why handing off the training stack matters.

The workload itself is shifting. Liu noted that most of Willow's work focuses on training, but inference is growing faster as dictation demand climbs — a shift that aligns with CoreWeave's broader inference push. The keyboard-free ambition Liu described for within two years follows from the same bet: that speech, cleaned up by a compact language model, becomes a practical alternative to typing.