MBZUAI Launches K2 Horizon 375B to Power AI Agents

The Institute of Foundation Models at Abu Dhabi's Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) has launched K2 Horizon 375B A23B. This flagship model leads a new six-model open-weights fleet released under the Apache 2.0 license. Moving away from the dense transformer architecture of its predecessor, the 70-billion-parameter K2 Think V2, the new flagship uses a sparse mixture-of-experts design. It features 375 billion total parameters with 23 billion active per token, alongside a massive 524,000-token context window. The model is available for deployment through partners including AWS, Cerebras, Nebius, and Compass.

K2 Horizon 375B A23B scored 47 on the Artificial Analysis Intelligence Index, representing a 30-point jump over K2 Think V2. In evaluations, the model prioritizes practical agentic workloads over trivia recall, outperforming its closest architectural rival, MiniMax-M3. On the GDPval-AA real-world knowledge work benchmark, K2 Horizon achieved an Elo rating of 1430 compared to MiniMax-M3's 1380. It also dominated the τ³-Banking tool-use simulation, scoring 34.2% against MiniMax-M3's 15.3%. However, it lagged on harder academic reasoning benchmarks, scoring 87.3% on GPQA Diamond and 32.0% on Humanity's Last Exam, compared to MiniMax-M3's 92.9% and 39.0% respectively.

A key feature for practitioners is the model's high rate of abstention on uncertain queries. On the AA-Omniscience benchmark, K2 Horizon attempted only 40% of the questions, helping lower its hallucination rate to 26% and raising its Omniscience Index from -40 to -3, despite a flat raw accuracy of 18%. Beyond the flagship, the K2 Horizon fleet includes a 0.9-billion-parameter model for wearable devices, 3.7-billion and 7-billion versions for mobile phones, and a dense 32-billion alongside a sparse 36-billion-A4B model for on-premise servers. This range allows developers to deploy tailored agentic capabilities across diverse hardware environments.