Abu Dhabi's MBZUAI Releases Six Fully Open AI Models, up to 375B Parameters

A full fleet, released open end to end

The Institute of Foundation Models, known as IFM, is the research arm of the Mohamed bin Zayed University of Artificial Intelligence in Abu Dhabi. On September 3 it released K2 Horizon, six models sized for very different jobs: a 0.9B model for wearables, 3.7B and 7B models for phones, a 32B dense model for local hosting, and two sparse mixture-of-experts models, a 36B that activates 4B parameters and a 375B flagship that activates 23B.

What separates the release from most open launches is how much came with it. IFM published the weights, the training code, the data-construction recipes or training data, checkpoints and evaluation information alongside the final models. That is the difference between shipping a car and shipping the factory that built it, and it is the claim on which IFM is staking the release.

Where it sits in the open field

The flagship 375B-A23B model is a reasoning model with a 524,000-token context window, and by activating only 23B of its 375B parameters at inference it aims to run far cheaper than its total size suggests. On the Artificial Analysis Intelligence Index it scores 47, against a median of 29 for open-weight models of similar size. IFM also points to two efficiency techniques, a diffusion-distillation method it says roughly triples generation speed, and a mixture-of-value-attention design meant to sharpen reasoning.

The release lands in an open ecosystem that has grown crowded and competitive, from Google's Gemma models passing a billion downloads to a wave of open-weight coding models reaching the frontier and Chinese labs like DeepSeek pushing agentic coding toward commodity pricing. K2 Horizon is the Gulf's bid to compete on openness itself, and by handing over the training data as well as the weights, it sets a higher bar for what other labs will have to call open.