Institute of Foundation Models

AI · Open-Source · Foundation Models

Launch

MBZUAI's IFM launches K2 Horizon, a six-model fully open AI fleet

Publishing full training data and methodology, not just weights, sets a new bar for what labs must disclose to call a model open.

• The Institute of Foundation Models (IFM) at MBZUAI released K2 Horizon on September 3, 2026: six AI models from 0.9B to 375B parameters, each shipped with weights, code, training data and methodology under Apache 2.0.
• IFM, founded by Eric Xing at Abu Dhabi's Mohamed bin Zayed University of Artificial Intelligence in May 2025, also runs research hubs in Silicon Valley and Paris.
• The six models share a common architecture and tooling but target different hardware budgets, from watches and phones (0.9B, 3.7B, 7B) to on-premise servers (32B, 36B-A4B) and enterprise reasoning (375B-A23B).
• The 7B model posted 70.6 on SWE-bench Verified versus 50.8 for the strongest comparable-size reference model, a notable jump in single-GPU coding-agent performance.
• The openness is uneven in practice: smaller models ship with data and code today, while the 375B and 36B-A4B cards still mark training data and intermediate checkpoints as forthcoming.
• IFM published a reward-hacking audit that lowered its own Terminal-Bench score from 70.2 to 66.9, an unusual public self-correction for a model vendor.
• As open-weight models jumped from about 11% to 62% of token volume on Vercel's AI Gateway between April and August 2026, enterprises are routing real traffic through models they can't fully inspect, raising the stakes for full-lifecycle transparency like IFM's.