# Milliwatt Labs

> AI measured in milliwatts, not megawatts. Milliwatt Labs is an independent lab building low-power, on-device AI: small models that run offline on everyday hardware. A sustainable AI future.

**AI measured in milliwatts, not megawatts.** We build small, efficient models that run on the devices you already own (phones, laptops, single-board computers and microcontrollers), with no data center and no GPU. Less power, more privacy, and AI that works anywhere.

What we build:

- **Voice**: High-quality speech in a tiny model for any CPU (Available) [Page](https://milliwattlabs.com/models/leokani-tts/)
- **Entertainment**: Songs with vocals, no data center needed (Available) [Page](https://milliwattlabs.com/models/yue2-3b-mlx-4bit/)
- **Chat**: Assistants that run on your device (Private beta)
- **Tiny devices**: AI on microcontrollers (ESP32) (Private beta)
- **Listening**: On-device speech recognition (In development)

Progress so far:

- 7M parameters ([Leokani-tts 7M](https://milliwattlabs.com/models/leokani-tts/))
- 20× real time on a $50 phone ([Leokani-tts 7M](https://milliwattlabs.com/models/leokani-tts/))
- Faster than real time on a Raspberry Pi ([Leokani-tts 7M](https://milliwattlabs.com/models/leokani-tts/))
- No GPU ([Leokani-tts 7M](https://milliwattlabs.com/models/leokani-tts/))

## Models

- [Leokani-tts 4.5M](https://milliwattlabs.com/models/leokani-tts-4-5m/): The same voice in a smaller model, for slower CPUs and microcontrollers. Just 4.3 million parameters.
- [Leokani-tts 7M](https://milliwattlabs.com/models/leokani-tts/): High-quality speech on any CPU, no GPU needed, in just 7 million parameters.
- [YuE2-3B 4-bit](https://milliwattlabs.com/models/yue2-3b-mlx-4bit/): Unofficial 4-bit MLX conversion of YuE2-3B by m-a-p: full songs with vocals from lyrics, on Apple silicon including iPhone.

## Blog

- [Leokani-tts 4.5M, and a second voice](https://milliwattlabs.com/blog/leokani-tts-4-5m-and-pono/): Leokani's voice in a smaller model, and Pono, a male voice. (2026-11-05)
- [Introducing Leokani-tts](https://milliwattlabs.com/blog/introducing-leokani-tts/): High-quality speech in a tiny model for any CPU, no GPU needed. (2026-10-29)

every release in full, with specs, tables, downloads and licence: https://milliwattlabs.com/llms-full.txt