Woof: A 32 MB text-to-speech model that runs offline, right in your browser
US startup Underdog has unveiled Woof, a text-to-speech model that weighs in at about 32 MB and works without an internet connection. A small retro browser game with a typewriter and a CRT monitor shows what it can do. The weights are on Hugging Face, but they're only free for non-commercial use.
What it's about
US startup Underdog unveiled a new speech model called Woof on November 6, 2026. Woof is a text-to-speech model: it turns written text into spoken audio. What makes it stand out is its size. The model is only about 32 MB, has around 31 million parameters, and according to Underdog runs on most devices with no internet connection at all. For comparison, today's large AI models have tens to hundreds of billions of parameters, making them thousands of times bigger.
The weights are available for download on Hugging Face. However, according to the Ukrainian tech outlet dev.ua, the license only allows non-commercial use. If you want to use Woof in a product, you need separate permission. TechCrunch was first to report the news.
The demo: a retro room in your browser
Instead of a dry spec sheet, Underdog built a small browser game to show off what Woof can do. You step into a virtual room full of old tech: a typewriter, a desk lamp, a CRT monitor and a computer. The lamp clicks when you hit the switch, and you can interact with many of the objects.
The two main attractions:
- The typewriter: Type any text on your keyboard and the game reads it aloud, generated by Woof, locally on your machine.
- The music tuner: Turn a dial and the system generates different kinds of music. According to dev.ua, the tracks you hear were created with Woof.
According to TechCrunch, the game's intro warns of "pushy" cloud providers in town, a not-so-subtle dig at AI services that run everything through someone else's servers. Underdog says everything in the demo runs directly in the browser using its own speech model.
How small is "small"?
Industry site FourWeekMBA pits Woof directly against the well-known open-source speech model Kokoro, which has 82 million parameters, and calls it an "8x smaller" model. Take that figure with a grain of salt: by parameter count, Woof (31M) is about 2.6 times smaller than Kokoro (82M). A bigger factor only appears if you compare file sizes, and those depend heavily on the number format the weights are stored in. There are no independent comparisons of audio quality between the two models yet.
| Model | Maker | Parameters | Size | Task |
|---|---|---|---|---|
| Woof | Underdog | approx. 31M | approx. 32 MB | Text → speech (and music in the demo) |
| Bark (Underdog) | Underdog | n/a | 167 MB | Speech → text |
| Kokoro | hexgrad | 82M | – | Text → speech |
Important: according to TechCrunch, Underdog's "Bark" is a speech recognition (speech-to-text) model for transcription and dictation. It has nothing to do with the older text-to-speech model of the same name from music AI company Suno.
What a model this tiny is good for
All sources agree: a model this size won't replace the big voice AIs. Expressive, emotional voices or cloning arbitrary voices are unlikely at 31 million parameters, and Underdog itself says nothing on the subject. For certain tasks, though, it may be enough:
- reading out news or notifications,
- generating personalized podcasts,
- powering voice assistants that shouldn't need to be online all the time.
The advantage of small models lies in what's called edge AI: AI that runs directly on your device (phone, laptop) instead of in a data center. That saves server costs, cuts latency and keeps data on your own device. Traders Union emphasizes exactly this point: less infrastructure, faster responses and less dependence on remote servers.
The company behind it
Founder Sigil Wen told TechCrunch he got started because, rather than using AI note-taking apps like Granola, he wanted to build his own system, with models that run on users' devices so they own their data. He has published a manifesto on the topic. He told TechCrunch:
"I believe we can run most of today's powerful models on the devices you already own within a few months. We just want to train the best open models: made in America, open and capable, and enabling private AI for people."
Woof isn't the only product. A few weeks earlier, Underdog released the Bark speech recognition model mentioned above (167 MB). There's also a 27-billion-parameter model designed to run on laptops, plus an app with a waitlist that Underdog positions as a locally run alternative to the AI services Muse and Instinct. The sources don't give details on these products.
What you need to try it
The easiest option is the browser demo: no installation, just a current browser. According to Underdog, Woof runs on most devices. If you want to download the model yourself, you'll find it on Hugging Face. At 31 million parameters and 32 MB, an ordinary computer is plenty. You don't need a graphics card; a modern processor and a few hundred MB of free RAM should be enough, and Macs and lower-powered laptops should work too. The available reports don't explain how to use the model outside the demo (software library, supported languages, voices), so reading the model description on Hugging Face is a must. Whether Woof supports languages other than English is unclear; the demo is in English.
The bigger picture
Woof isn't a breakthrough in audio quality. It's a statement about size: speech output that fits in 32 MB and runs offline in your browser is interesting for tinkerers, app developers and anyone who doesn't want to send text to cloud services. The playful demo makes the idea tangible. Some caveats remain: there are no independent tests of sound, speed or language support yet, and the "8x smaller than Kokoro" claim doesn't hold up on parameters. "Open" is also limited here. The non-commercial license sets Woof clearly apart from Kokoro, which uses the permissive Apache 2.0 license. If you want to build a product, you'll have to negotiate with Underdog. The interesting question is whether Underdog delivers on its promise to bring much larger models to everyday devices.
Sources
- Google News (US) dev.ua
- Google News (US) fourweekmba.com
- techcrunch.com techcrunch.com
- tradersunion.com tradersunion.com
Try it yourself: Woof
Explained step by step, no prior knowledge needed – on your own computer or free in your browser.
What can I do with it?
Woof reads typed text aloud, entirely on your own device, with nothing sent to the internet. At only about 32 MB, it's small enough to run right in your browser. In Underdog's demo, the system also generates various music styles via a virtual radio tuner. Its strengths are its size, its speed on ordinary hardware and privacy. The limits: a 31-million-parameter model won't match big cloud voices on expression, emotion and naturalness, and there are no independent audio tests yet. Compared with the similarly small Kokoro (82M parameters), Woof is even more compact but has a more restrictive license. It's unclear whether languages other than English are supported.
Who is it worth it for?
Worth it for tinkerers, developers of offline apps, browser games or voice assistants, and anyone who doesn't want to send text to cloud services. If you're building commercial products, you'll need permission from Underdog. If you expect studio quality or non-English voices, you're probably better off with larger models.
Ideas: What you can do with it
Have your own notes, emails or articles read aloud without the text ever leaving your device. Just enter the text, generate the audio and listen on the go.
Write yourself (or have a chat model write) a short daily summary with appointments and to-dos, and let Woof voice it. The result is a mini podcast just for you.
As Underdog's demo shows, Woof fits inside a browser game. Game characters can speak typed or generated dialogue, with no server needed.
For a homemade voice assistant or smart home project, Woof can announce status messages like "The washing machine is done." That works without internet, too.
Developers can use Woof to quickly test how speech output feels in an app. For commercial use, though, you'll need a license from Underdog.
What your computer needs
According to the sources, there's only one variant, with about 31 million parameters and roughly 32 MB. It's small enough for practically any computer, so no shrinking (quantization) is needed. To get a first impression, the browser demo is all you need.
- Model size of about 32 MB: it fits on practically any device.
- No graphics card required. According to Underdog, Woof runs on most devices, even fully offline.
- The demo runs directly in your browser; a current browser is all you need.
- The model should run fine on Macs and lower-powered laptops too. A few hundred MB of free RAM is enough.
How do I find out what my computer has?
- Windows: Press Ctrl + Shift + Esc → Task Manager → Performance tab → click GPU on the left. At the bottom you'll see Dedicated GPU memory (that's your VRAM). You'll find system memory under Memory.
- Mac: Apple menu at the top left → About This Mac. You'll see your chip (e.g. M2) and memory (e.g. 16 GB). On Apple chips, graphics and processor share this memory.
- Linux: Type
nvidia-smiin the terminal (shows your graphics card and its memory) andfree -hfor system memory.
Guide
This model runs with its developers' own program. The commands below come from the Model Card and the developers' instructions.
Windows 10 / 11
Windows 10 / 11: not supported. Woof is only available in MLX format. MLX is made for Macs with Apple chips, and the source doesn't mention a Windows version. At most, you can try it at your own risk via WSL2 (Linux running inside Windows), as described in the Linux instructions. That hasn't been tested, though.
macOS
macOS: supported. Woof comes in MLX format. MLX is Apple's compute library for Macs with Apple chips (M1, M2, M3, M4 …). It won't run on older Macs with Intel processors.
- Check the prerequisites. You need a Mac with an Apple chip (M1 or newer) and at least 8 GB of RAM. You also need Python 3 and Apple's developer tools. Open the Terminal app and enter the command below. If a window pops up, confirm the installation. Then run "python3 --version" to check that Python is installed.
xcode-select --install python3 --version - Create a working folder. Create a dedicated folder for Woof and change into it. Everything else happens in this folder.
mkdir woof cd woof - Create a virtual environment. A virtual environment (venv) is an isolated space for Python packages, so nothing clashes with other programs. Once activated, "(.venv)" appears at the start of the line.
python3 -m venv .venv source .venv/bin/activate - Install mlx-lm. According to the Model Card, mlx-lm is the tool used to load Woof. Installation takes a minute or two. If there's no red error at the end, it worked.
pip install mlx-lm - Download the model. This command downloads Woof from Hugging Face (a platform for AI models). It's about 3 GB. If the "hf" command isn't found, install it first with "pip install huggingface_hub".
hf download ConwayResearch/woof-1.0-4B - Generate your first reply. Now ask Woof a question. The mlx_lm.generate command is part of mlx-lm and prints the reply straight to the terminal. You can change the question between the quotation marks to anything you like.
mlx_lm.generate --model ConwayResearch/woof-1.0-4B --prompt "Hello! Please briefly introduce yourself." - Result. The terminal shows the text reply Woof writes in response to your question.
Linux
Linux: not officially tested. The Model Card only mentions mlx-lm and mlx-swift, meaning Apple's MLX. The source doesn't say whether MLX runs on Linux with a CPU or NVIDIA graphics card. These steps are therefore an untested attempt. Woof is only guaranteed to work on a Mac with an Apple chip.
- Check the prerequisites. You need Python 3 and at least 8 GB of RAM. Open a terminal and check the Python version. If a version number appears, Python is installed.
python3 --version - Create a working folder. Create a dedicated folder for Woof and change into it.
mkdir woof cd woof - Create a virtual environment. A virtual environment (venv) keeps Woof's Python packages separate from the rest of your system. Once activated, "(.venv)" appears at the start of the line.
python3 -m venv .venv source .venv/bin/activate - Install mlx-lm. Install mlx-lm, the tool named in the Model Card. If the installation fails with an error, or the model won't load in the last step, MLX isn't supported on your system.
pip install mlx-lm - Download the model. This downloads Woof from Hugging Face, about 3 GB. If the "hf" command is missing, install it first with "pip install huggingface_hub".
hf download ConwayResearch/woof-1.0-4B - Generate your first reply. Ask Woof a question. If everything works, the reply appears in the terminal.
mlx_lm.generate --model ConwayResearch/woof-1.0-4B --prompt "Hello! Please briefly introduce yourself." - Result. The terminal shows the text reply Woof writes in response to your question.
What to do after your first result
- Change the text after --prompt and ask Woof your own questions.
- The model is licensed under CC BY-NC 4.0. You may only use it for non-commercial purposes and must credit Underdog as the source. For commercial use, you need a separate license, which you can request at license@conway.tech.
- According to the Model Card, Woof can also be loaded with mlx-swift. That lets you build it into your own apps for Mac or iPhone.
Try it
Hello, this is a short test of a tiny speech model running on my own computer.Good morning! Today is Friday, the weather is sunny, and you have three meetings.Welcome to my podcast. In today's episode we talk about why small AI models matter.Wait... did you hear that? Someone is knocking at the door!The quick brown fox jumps over the lazy dog, 1234 times, at 5:30 p.m. on November 6th.Tips for better results
- Start with the browser demo to judge sound and pacing before you download the model.
- Test short, clearly structured sentences with punctuation. Commas and periods help small models with emphasis and pauses.
- Spell out numbers, times and abbreviations if they're read out incorrectly.
- Split long texts into paragraphs, voice them one at a time, then stitch them together.
If something goes wrong
Mind the license: without separate permission, only non-commercial use is allowed.
Don't mix them up: Underdog's Bark is a speech recognition (speech-to-text) model, not Suno's text-to-speech model of the same name.
How to use the model outside the demo is explained on the Hugging Face model page, so read the description first.
Non-English text may not be pronounced correctly; the demo is in English.
Woof isn't in the Ollama library. Ollama is meant for text chat models anyway, not speech output.
“out of memory,” a crash, or your computer freezes
The model is too big for your VRAM (video memory). Pick a smaller variant from the table above and close other programs, especially games and browsers with lots of tabs.
It runs, but extremely slowly
Your CPU is probably doing the work instead of your graphics card. In Ollama, the PROCESSOR column of ollama ps shows how much runs on the GPU. If it says "100% CPU", update your graphics driver or pick a smaller variant that fits entirely into VRAM (video memory).
“ollama” or “python” is not found
Close the terminal window and open it again so it recognizes the newly installed program. For Python on Windows: reinstall it and check Add python.exe to PATH.
Windows: "Running scripts is disabled on this system"
Run the script with exactly the command above: powershell -ExecutionPolicy Bypass -File try_windows.ps1. This allows execution for this one time only.
Error message when loading the model (version, "trust_remote_code", unknown architecture)
Brand-new models often need the latest software version. With Ollama: update Ollama. With Python: pip install -U transformers in the activated environment. The Model Card usually lists the exact version.
Tokenstrom generates the scripts and notebook automatically. Please take a quick look before running them – they're plain, readable text. Mind the model's license: cc-by-nc-4.0.
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.