Before you paste it into AI, ask who it answers to. Head of R&D @TrellisData, the company behind @AgentausTD: frontier AI that never trains on your data.
@unleaded_dev@unleaded_dev In the meantime, a few results: MMLU-Pro 0.844, MMLU 0.872, LegalBench 0.807, DocFinQA 0.623, RULER 128k 0.899, BFCL 0.922, and LongBench v2 0.664 on documents up to ~980k tokens. Single runs, about Β±2 points.
Ten years of R&D in Canberra, Australia's bush capital, and the number I'm proudest of is 2.
Agentaus 1 is a 220B model that runs on half the GPUs of similar-sized models: 2 instead of 4. That's what keeps it fast, affordable and better for the environment.
Engineers: what would you build with it?
https://t.co/wniXuFtnxq
We built Australiaβs own AI model, and today itβs open to the world.
Agentaus 1 (@AgentausTD): 220B parameters. 84.4% on MMLU-Pro. Ranked 43rd in the world.
It runs on Australian servers. Nothing ever leaves Australia, and nothing you type is used to train AI. Delete a chat and itβs gone for good.
Most of the worldβs AI answers to somebody else. This one answers only to you.
Free to start: https://t.co/pkgBYQlpkZ
@sunkcostcapital@sunkcostcapital On the "limited detail" point, here's more: MMLU-Pro 0.844, LegalBench 0.807, RULER 128k 0.899, BFCL 0.922. Single runs, so Β±2 points or so.
@unleaded_dev@unleaded_dev Thanks! No white paper yet. It's our own proprietary architecture, different from the existing open-source models on the market.