Med42 v2

Med42-v2 is M42 Health's suite of open-access clinical large language models, instruct- and preference-tuned from Llama-3 (8B and 70B) on a curated medical dataset built with Cerebras on the Condor Galaxy supercomputer and evaluated with MBZUAI. The 70B variant reported 79.1% zero-shot MedQA — state-of-the-art among open medical LLMs at release — and topped the Clinical Elo leaderboard ahead of GPT-4-class models.

It is positioned as clinical decision-support infrastructure for the UAE's health AI ambitions, with weights on Hugging Face (gated) and broad quant availability (GGUF, AWQ).

Honest limits: gated behind a custom M42 license (not a standard open license); the authors stress it is not ready for real clinical use — extensive human evaluation ongoing, with risks of incorrect/harmful outputs and training-data bias. Never use for actual medical decisions without rigorous safety testing.