Epoch AI Labs

@KridayTop Bro I'm tired 😭

Guys get ready : Something cool is coming up from OUR side @EpochAILab

Hello 👋 , after the release of playbox - we aren't stopping. Next up is the model named Gauss named after the prince of mathematics Carl Friedrich Gauss. This model should be available in the coming months as open weights. Read more here : theepoch.in/pages/gauss-fu…

Feeling cute. Might drop a model later 😜

So we routed instead — kept the base model untouched, added a separate router for tool calls. Result: 50.3% GSM8K (beats Qwen2.5-3B's 49.0%, at 1/4 the params) and 75.5% on routed OOD arithmetic. Full writeup + what we tried before this worked: theepoch.in/pages/playbox-…

We tried fine-tuning tool use directly into a 0.8B model. It broke general behavior. The model got worse at everything else. Thread(🧵)

Check out the full story on : theepoch.in/pages/playbox-… This is our first model , Hope You all like it !

Today we have @EpochAILab have posted our first model of many called Playbox 🧰. Here is the break down (🧵thread ) We started with Qwen 3.5 0.8B , It already had great mathematical prowess but lacked the ability to actually be useful in math.

@KridayTop 🎊🎊 ?!

We have posted a blog on playbox release. Have a good read friends 📑! theepoch.in/pages/playbox-…

Shipped Playbox . 0.8B params. Built for math reasoning + routed tool use. 50.3% GSM8K neck-and-neck with Qwen2.5-3B at 49.0% 75.5% on routed OOD arithmetic Small model. Real benchmark. Close race. 📦

@KridayTop 👀

Playbox is a test model on maths and will be the basis of our other models from @EpochAILab along with our research paper . Final version for playbox aims to achieve 60%+ on evals we are doing and then full set evals. Paving the way for the other 3 giants coming soon! (For math)

Today I tested 3 models against each other in Model Kombat. Qwen 3.5 0.8B and 2 of my models (pre and post rft) . The results were surprising but not game breaking . Thread 🧵

We mapped behavioral circuits, froze parts of the model during training, and tested whether safety and reasoning could improve without erasing capabilities already learned. We’ll be sharing the results, methods, and failures here.

Our first research began with a simple question: When we change a model’s behavior, how much of the model actually needs to change? Structural Alignment is our attempt to find out.

We started Epoch because building more capable models isn’t enough. We need to understand what changes inside them, what remains intact, and what we trade away. Our first research program is Structural Alignment. New models. A new epoch.