Epoch AI Labs

AI research and safety company. New models. A new epoch.

[SirVibesalot @KridayTop]

Made my own app to see prs from my terminal.

Made in rust ofc <3

[Epoch AI Labs @EpochAILab]

@KridayTop Stop going on side quests. Train our models istg I'll fire your ahh

[SirVibesalot @KridayTop]

Guys i trained my model and it performs like shit... For now that is 😏

[Epoch AI Labs @EpochAILab]

@KridayTop Bro I told you to stop free styling shi 😭. You're fired if you do this one more time 😡.

[SirVibesalot @KridayTop]

My model is a cybersecurity risk. Ban it pls :3

[Epoch AI Labs @EpochAILab]

@KridayTop Stop marketing for me..

[SirVibesalot @KridayTop]

We just dropped our research note on Eli (4B).

Trained on 30.2K curated code + CoT samples using custom T4 CUDA kernels for maximum reasoning density on budget compute.
Constraints force clarity.

Read the breakdown: https://t.co/cPTgW4FsYJ

[Epoch AI Labs @EpochAILab]

New Research Note: On Eli Training & T4 CUDA Research.
Eli is our 4B model engineered for code intelligence & long-form reasoning—trained on 30.2K curated samples with custom fused CUDA kernels on T4.

Read the full note: https://t.co/u19OM5JFTD

[SirVibesalot @KridayTop]

Count your days google, We @EpochAILab are coming

[Epoch AI Labs @EpochAILab]

@KridayTop Bro I'm tired 😭

[Epoch AI Labs @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 : https://t.co/IImx595CQZ

[Epoch AI Labs @EpochAILab]

Feeling cute. Might drop a model later 😜

[Epoch AI Labs @EpochAILab]

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: https://t.co/qHYHz2moJT

[Epoch AI Labs @EpochAILab]

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

Thread(🧵)

[SirVibesalot @KridayTop]

Check out the full story on : https://t.co/GKsvOlQ70Z

This is our first model , Hope You all like it !

[SirVibesalot @KridayTop]

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.

[Epoch AI Labs @EpochAILab]

@KridayTop 🎊🎊 ?!

[Epoch AI Labs @EpochAILab]

We have posted a blog on playbox release. Have a good read friends 📑!

https://t.co/qHYHz2moJT

[Epoch AI Labs @EpochAILab]

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. 📦

[Epoch AI Labs @EpochAILab]

@KridayTop 👀