Engram runs a tiny offline audio model

On September 27, 2026, The Verge reported that Thoughtful Things opened a Kickstarter for its first instrument, Engram. It is a sampler and groovebox. It uses AI to mangle audio coming in, and it can also hallucinate new sounds. The report says this is not Suno in a box, and not a button that returns a song ready for pop radio. The aim is experimental, uncanny sound, and pushing audio models past the point where they stay stable.

Engram is not on the internet. It runs a tiny AI locally. The Kickstarter listing, as reported, says the model is designed in-house and custom-trained.

The Verge quotes the company: the audio models were trained on open datasets that contain only audio licensed for commercial use, such as CC-BY or similar. The company says it has not trained, and will never train, on non-commercial, pirated, or otherwise stolen data.

The company also plans to open the firmware, so other people can change it or load their own models. In the campaign video, founder Evan King uses his voice to ask for piano, and the machine returns something glitchy and only vaguely piano-like. The company's image for it is a field recorder for latent space. It also borrows from circuit bending: the small model can be tweaked, and it can be broken on purpose.

The Kickstarter is a limited run, starting at $675. The company has not stated a retail price. The Verge says the retail price is likely between $850 and $900, because the $675 pledge is listed as a 30 percent discount. That range is the report's inference from the discount, not a price the company published.

The training-license claim is the company's own statement. The article does not name the datasets, and it does not cite an outside check. Opening the firmware is still a plan. This piece did not open the Kickstarter page or play the demo.