On 3 September, MBZUAI's Institute of Foundation Models (IFM) released K2 Horizon: a fleet of six fully open-source AI foundation models designed for advanced reasoning, mathematics, coding and complex agentic tasks, shipped with their code, training data and methodology. The flagship is a 375-billion-parameter model built for enterprise reasoning, with five smaller models spanning everything from wearables and phones to local hosting.
Open source most often means the weights alone: the data and training method behind a model usually stay private. IFM has published all three, the weights, the data and the method, so the results can be reproduced and checked rather than taken on trust. For the UAE, which is building its AI strategy around enabling sovereignty, this is a crucial distinction: a model customers can inspect, reproduce and adapt in-country is a fundamentally different proposition from one they can only access as a black box, however capable that black box is. Core42's sovereign enabled infrastructure is where that fleet becomes something enterprise and government customers can deploy. Through Compass, they get K2 Horizon inside the same sovereign enabled environment, with the same data residency and security controls, they already run their other AI workloads in. In doing so, Compass reinforces the UAE's broader strategy of building sovereign enabled, fully reproducible frontier AI systems through national institutions and making them openly available for wider use. Customers are responsible for assessing applicable regulatory, sector and data-classification requirements and obtaining any required approvals for their use case. The solution is ideal for sensitive data but not secret or classified data, which should leverage fully private cloud environments.
What that looks like in practice
Consider a bank's security team piloting a fraud-detection workflow that can't let data leave the country, or a government agency that needs to review a model's provenance, evaluation evidence and deployment controls before letting it anywhere near a decision involving sensitive information. Today, that typically means either building the infrastructure in-house or working within whatever a vendor happens to provide. With K2 Horizon on Compass, these teams can run a frontier-class model entirely inside a sovereign-enabled environment, with the training data and method published clearly enough to review if they need to.
Six models, built for different deployment contexts
The fleet spans six sizes, giving developers and organizations the flexibility to match the model to the deployment requirement. At the smaller end, the 0.9B model is designed for highly constrained environments such as watches and glasses, while the 3.7B and 7B models bring advanced capabilities to phones and other on-device applications, where inference needs to run close to where data is captured; all three set a new state of the art at their respective scales. Two mid-sized models, a dense 32B and a 36B mixture-of-experts with roughly 4B active parameters, support local and on-premise hosting for organizations that need the model to run entirely inside their own environment. The 375B flagship, a mixture-of-experts model with 23B active parameters built for enterprise reasoning, sits at the top of the range. All six models and their code are released under the Apache 2.0 license, clearing the way for commercial use.
Where it fits on Compass
Compass organizes its catalog around nine capability categories, including text generation, embeddings, audio, image, vision, function calling, real-time, batch and document AI, plus fine-tuning and integrations with LangChain and Azure AI Search. K2 Horizon's strengths point to where it fits most naturally on Compass. Its general-purpose reasoning makes it a strong fit for text generation, and its 512K-token context window suits high-volume work through the Batch API, ideal for organizations processing large document sets or running analysis outside a live chat session.
Why Compass
IFM named four inference partners for K2 Horizon at launch: Compass, Cerebras, AWS and Nebius. All four offer the same weights, so the difference for a customer isn't the model, but the platform around it. Compass makes K2 Horizon usable as part of an enterprise AI platform: UAE region access, documented APIs, model choice, usage management, security controls and the infrastructure required to move AI workloads from experimentation toward production. Sovereign enablement is what sets it apart. Compass runs on Core42's sovereign enabled infrastructure in the UAE, so customers access the model inside an environment that is built, operated and governed in-country rather than across borders. Two capabilities are particularly important for enterprise and government customers:
Data security. Compass does not persist user prompts or model responses. Prompt and response data remain transient for the duration of the inference request and are not written to persistent storage, reducing unnecessary exposure of sensitive enterprise data.
UAE data residency and in-country inference. For K2 Horizon deployed on UAE infrastructure, inference is performed within the UAE, so customer data remains within national boundaries rather than being routed to overseas infrastructure. Compass provides the underlying geo-restriction, encryption and access-control framework required to support UAE data-residency requirements.
Open models are already being leveraged beyond research. The next challenge is making them usable inside the trusted environments that enterprises and governments already run. This is where Compass plays a critical role, providing a sovereign-enabled, enterprise-grade environment through which organizations can access and deploy K2 Horizon with the security, control and data residency requirements they need.