Companies around the world, including those labeled AI, are competing to create larger and better models. However, the problem is that these models require a substantial amount of computing power. On the other hand, some researchers and startups are developing smaller models that can run on edge devices such as smartphones or computers.

One of the startups, named Underdog, released a game created using its new voice model Woof, which is only 32MB in size.

The scene of this game is a room with a retro-tech style, which contains a typewriter and a CRT monitor. There is also a table lamp in the room that makes a sound when its switch is clicked. At the beginning, the user is prompted to be aware of the “intrusive” cloud service providers in the city – which is clearly an allusion to cloud-based models.

You can interact with typewriters and music tuners.

You can type any content on the keyboard on the typewriter, and the game will read it out. You can also rotate the tuner to generate different types of music.

All of this is run in the browser through the voice generation model of Underdog. The company states that this model can run on most devices without any need for an internet connection. The model has 31 million parameters and is available for download to everyone.

Although such a small model is unlikely to be able to replace larger models, it may still be sufficient to drive the experience of local assistants or to be used for creating personalized podcasts.

A few weeks before the release of this demonstration, Underdog just launched Bark, which is a 167MB speech-to-text model also designed to run on edge devices for transcription and dictation purposes.

Underdog founder Sigil Wen told TechCrunch that the reason he started developing this application was because he didn't want to use a note-taking tool like Granola, but rather wanted to create his own system so that the AI model could run on users' devices and allow users to have their own data.

“I think that within a few months, we will be able to run most of the current strongest models on the devices you already have. We just want to train the best open-source models, which are made in the United States, open-source, powerful, and will allow people to have private AI,” he said, referring to his own declaration.

In addition to the two speech models mentioned above, Underdog has also released a model with 27 billion parameters that can run on laptops; moreover, the company also has an application that is waiting in line to be used, positioned as a local hosting alternative for Muse and Instinct.