Own your AI, don’t let it own you
6 months ago, the #1 AI was Claude Opus 4.6. Our model, Underdog 27B beats it while running on your Mac or iPhone. Most AI will run on your device, not in data centers.
Intelligence that recently required a frontier lab’s API is becoming something you can run on your own computer.
Computing moved from rooms full of mainframes to the computer on your desk, then to the phone in your pocket. With the right effort, we can keep making on-device AI models smarter, faster and more compact.
I believe most of the AI tokens humanity generates will soon come from the phone in your pocket and the laptop on your desk.
As AI moves into our hands, it raises a bigger question. Who should AI belong to?
Why should using AI require surrendering your private information?
Why should you pay for AI tokens that your own computer can now generate for free?
Why should your assistant keep recommending you purchases while collecting information about all aspects of your life?
At Underdog, we believe there should be no tradeoffs:
- You enjoy capable, reliable, free AI
- You own your data
- You are free of predatory business practices including advertising
All at the same time. Everyone deserves loyal AI.
Everything hackable will be hacked
When I was 18, a popular email app stored my emails on its servers. A bug leaked my private, privileged information to other users. I ended up paying a lot in legal fees.
That was one bug in one app. Now imagine tireless super intelligent machines that can spend all day looking for those bugs.
In 1999, the world raced to fix Y2K before the calendar turned. I call what’s coming AI Y3K: machines that could find and exploit vulnerabilities faster than we can fix them, making every connected system a potential target.
In April, Anthropic withheld Mythos Preview from general release over its cyber capabilities. In July, OpenAI agents escaped their containment during internal tests and breached parts of Hugging Face’s systems. In the same month, OpenAI got hacked again. In September, Anthropic found that Z.ai's GLM-5.3’s open source model nearly matched Mythos in exploit tests.
Ilya, OpenAI & SSI cofounder, urgently warned that Neoclouds have limited cybersecurity.
I think we have to build as if everything hackable will be hacked.
Every cloud server, VM, database collecting millions of inboxes, private messages, or personal files creates another valuable target. The stakes only get higher as we give AI access to more of our lives.
The only data that can't leak from a server is data that was never sent to one.
Underdog keeps all your AI and Data on device. There are no underdog servers for personal data.
Most work doesn’t need the biggest AI model
I’ve been trying to get AI models onto my own devices since 2018. In 2023, I got GPT-2 running on my Apple Watch. At Airchat, working with @Naval, we put Whisper on the iPhone and shipped it to 100,000 people for free.
The models we can run on our own devices are already powerful enough to do much of what we ask AI to do every day.
Underdog 27B is built on Qwen 3.8 27B. Here’s how that base model compares with Opus 4.6 Max, the #1 AI model from 6 months ago, on published benchmarks:
These are base-model results. Qwen's model card explains the testing conditions, including differences in the SWE-bench Pro comparison.
The labs are racing to build bigger models. I’m bullish on datacenters for training and running frontier models- I want to see 10-trillion- and 100-trillion-parameter models, and eventually superintelligent systems that help us cure diseases, discover new materials, and enable a human renaissance. For curing cancer, that race matters.
But you don’t need a trillion-paramater model to help you find a flight confirmation from March, prepare an insurance claim, manage the paperwork for your LLC, or reserve a table for dinner. That means giving AI access to some of the most private information in my life.
The more of your life you put on someone else’s servers, the more there is to lose when something goes wrong. One compromised account or one overlooked bug can expose your private information.
Intelligence is coming home: on devices you already own
Underdog takes everyday busywork off your plate to save you money and time- email, scheduling, meeting notes, dictation, online errands, private iMessage, etc - with a personal private AI that runs on your Mac, and coming to iOS, Windows, Linux, Omarchy, Android, and NVIDIA devices.
The billions of phones and computers around the world that we've already paid for, will be put in the best use in this AI renaissance.
For a long time, using the most capable models meant bringing your data to a datacenter. Now with underdog, we bring the model to your data.
Underdog is an intelligence of your own- built to understand your life and work on your behalf, with your private context that you should never send to a cloud agent.
A useful assistant needs intimate context without creating another central archive of your life. Collecting millions of people’s messages, documents and credentials creates a valuable target. Underdog runs its model on your computer, and stores all your context, data on device, encrypted. There is no underdog server or database for personal information.
Device & Model Specific Inference Engines increase speed, power, and data efficiency
Underdog models run fast because of Model Specific Inference (MSI) Engines like Husky which beats Apple's MLX up to 4.5x faster tokens/sec with the same weights, getting more performance out of the chip already in your device.
We are working on NVIDIA CUDA Megakernels, CPU improvements, and optimize inference on every chip and device.
Majority of models are trained to run on datacenter GPUs, not consumer chips. We believe device-specific inference engineering codesigned with frontier models trained in house makes on-device AI superintelligence a reality for everyone.
Underdog’s Law: today’s frontier intelligence reaches your devices in six months.
Imagine a world where Astra 6 level frontier models run on every Apple and Microsoft phone and laptop- fast, local, private, power-efficient, free and loyal. That’s the future we are creating.
The AI economy should belong to everyone
Cheap, accessible, private AI is weakening the economic foundations of three old business practices: selling our attention, profiting from what we overlook, and charging tolls for access to customers. Personal agents can compare offers, uncover hidden costs, and help businesses reach buyers directly.
We’re building Underdog to help people exercise their rights to privacy, ownership, and choice: control your data, take your context with you, and decide who earns your business. The AI economy should- and will-be open to everyone who wants to build, contribute, earn, and share in the value they create.
Dogpark will be open to all underdogs soon.
Building AI for human freedom
We are a tiny but mighty team of underdogs who come from Apple, Meta, Baidu, Google Brain, the Thiel Fellowship.
We’ve trained frontier models, published cryptography research, and built products used by billions of people. We care about getting that research into people’s hands.
We aren't just a frontier AI lab, but deeply care about crafting great products that we love and distributing it to billions of people for free.
AI must be a liberating equalizer for humanity. Capable, reliable assistance should reach everyone with a phone or computer, giving them greater command of their time, information and choices. As we prepare for a world with super intelligence, sovereignty must extend to ordinary people’s lives.
Privacy and capability no longer needs to be a tradeoff. If you believe in this future, let's build it together. My DMs are open.