Not sure if this has been shared already, but AllenAI / Ai2 is a US-based nonprofit who are trying to build AIs as open-source and transparently as possible.
Their OLMO models have fully transparent training data. Their Tulu ones are as transparent as you can be building on top of Llama. For some positive news out of the US this week, they released their new 405B Parameter model for free online chat and download.
Chat: https://playground.allenai.org/
HuggingFace: https://huggingface.co/allenai/Llama-3.1-Tulu-3-405B
Is it any good?
There are benchmarks on the huggingface page. The larger model is close to GPT4o performance. Which makes this worse than deepseek-r1. But it is a smaller model and not a reasoning model (doesn’t use up extra tokens to “think”). So still very impressive and important for open source.
