Meet Thud: the first Audio Action Model.

Meet Thud: the first Audio Action Model. It listens, chooses an action, and executes it without generating a transcript first. Think of it as a "Jev" that can hear you.

Open an app. Navigate a menu. Send a message. Just say it.

In our tests, roughly 10× faster and 20× cheaper than Gemini Live.

Watch me navigate my smart glasses at <300ms latency, just by talking to Thud.

Thud is the first model of its kind: it listens to audio and acts on it 10x faster and 20x cheaper than Gemini Live. It goes directly from audio → action. That means it doesn’t try to “write down” what you said and instead just does it. It can run directly on your phone.

Here’s what we built in our first day with Thud.

• Smart glasses you can navigate just by talking to it. It’ll listen, classify the action you say, then do it super quickly. This works for navigating on smart glasses and even within specific phone apps. 

• Fixed Spotify’s search. Thud can search your Spotify by audio.

• An AI that listens 24/7, and actually knows when you’re talking to it vs somebody else. Based on what you’re saying, how you say it, and the surrounding audio context, it figures out if it’s worth jumping in and saying something…
• Thud will even route you to the right model based on what it hears. It knows whether you’re talking to your Claude Code, your Muse, or your ChatGPT, or whether to stay quiet if you’re just chit-chatting.
• An alarm that warns you when the faucet is left on, the doorbell rings, or an ambulance passes by. Thud hears it and it knows.

Why audio action models are different

Every app tries to “solve audio” by turning it into text. But that doesn’t work, because you miss all of the human emotion, inflection, and sounds. And while LLMs are cool, it’s lame to flatten an audio problem into a text problem. That’s why Audio Action Models are so cool.

I also believe that not every problem is a “text output” problem. If you can reshape your problem into a yes/no answer (i.e. “What did the speaker say to do?” → “Did the speaker say click this?”), you can make things smarter and faster. This was shown recently, with Jev, which “solves” text classification.

But this hasn’t really been done with audio before. Here’s what happened when we brought the lessons from Jev to the world of audio.

• It classifies in 107.6ms and runs locally.
• It gives confidence probabilities for each category.
• Most models can only recognize people talking. But Thud can hear weird things like “a leaky faucet”, “a young man laughing heartily”, and “two people snapping their fingers while singing”. 
• No hallucinations: it never categorizes outside the given choices.
• You can change the categories as easily as changing a prompt.
• It’s audio native: there’s no intermediate speech-to-text layer like most systems.
• Thud doesn’t care about order. With a regular LLM, if you give it options to choose from, the order in which you give the items will determine what it categorizes. That’s terrible. This model fundamentally doesn’t care about order.

Smart glasses need local models to work.

When you’re talking to your glasses, the response needs to be near-instant, it needs to be cheap enough to run all the time, and it needs to be private.

Proprietary models can’t achieve any of this. Gemini Live is $0.005/min. Using it for 10 hours/day would cost $90/month. OpenAI’s gpt-live would be $900/month. As time goes on, cloud models will (hopefully) get cheaper, but they certainly won’t get more private. You are sending every word you say to Google. Lastly, cloud models aren’t instant. It is (virtually) impossible to achieve 100ms latency for a cloud model. But this is possible with local models. This is possible with Thud.