A closer look at what we've actually built inside BerryGPU.
The whole thing runs as a little operating system in your browser. You have a desktop with draggable windows, a task manager, a wallet, and a bunch of applications that actually do something.
miner.exe uses your GPU to run SHA-256 hashing through WebGPU. Every valid share gets verified by our coordinator and credited to your wallet. Every 2³⁴ hashes earns you one compute credit.
train.exe lets you train AI models using those credits. You can upload your own datasets, use Q&A pairs or text, and watch the training loss and performance as the run progresses. There's also local training that runs directly in your browser for free.
network shows our fleet of 1,048,576 virtual GPU addresses. Training jobs are broken into smaller pieces, routed through the fleet, and calculated on real CPUs. Every run verifies the combined result against a direct calculation.
models is where everything you've trained ends up. You can chat with your models, publish them, and let other people try them. There's also a community library to explore what others have made.
hermes lets you connect Hermes Agent from Nous Research to the BerryGPU environment. We're interested in giving agents access to the same compute and training tools that people use.
We've also built a live task manager showing mining performance, a leaderboard for miners, a hardware benchmarking app, dataset management, and a wallet that works without passwords or gas fees for signing in.
Every network training run also gets its own public receipt with the loss curve and verification results, so people can inspect what actually happened.
All of this runs from one browser tab. We're trying to make compute something people can actually use, experiment with, and understand rather than just another service they rent.
https://t.co/uBzTOQc5X6
BerryGPU is pretty simple. Open your browser, let your GPU earn compute credits, use those credits to train an AI model on your own data, and then chat with it or share it with others. You can even train smaller models locally for free. No cloud setup, no credit card, just your computer.
Training your own AI model shouldn't mean renting expensive GPUs, setting up cloud accounts, or spending hours figuring everything out. That's one of the main reasons we're building BerryGPU. We want anyone to be able to open their browser and start experimenting.
@_deejaycryp Appreciate it man, would love to hear what you have in mind. We're still pretty early with BerryGPU and definitely open to ideas, especially anything that gets more developers actually building with it.
BerryGPU is live. We've built an entire computer inside your browser where you can mine compute credits, train your own AI models, and watch everything happen in real time. There's also a network of over a million virtual GPU addresses handling distributed training workloads. We're now working on bringing Hermes into the system so agents can use that compute themselves.
$BERRY: 9nrPM6kGijkrPn6CAeWHZbRrvVKq5gmxQPdC4zMVpump
@_deejaycryp I think the best way to bring people in is to just let them use it. With Hermes we're making it possible for developers to connect their own agents and put the compute to work. We'd rather have people building and experimenting with it than spend all day trying to convince them.
Hey @Teknium we've been building something called BerryGPU. It's basically a little operating system inside your browser where anyone can use their GPU to mine compute credits, train their own AI models, and access a network of 1,048,576 virtual GPU addresses for distributed training.
We're now looking at bringing Hermes into the system so agents can manage their own compute, run training experiments, use tools, and improve through actual work.
Would love to get your thoughts on this. Think Hermes could be a really interesting fit here.
https://t.co/uBzTOQc5X6