AIBRIEF
The K2 Horizon family runs from a small model for smartwatches and other low power devices to a 375 billion parameter system for large organizations.
Abu Dhabi's IFM released six AI models on Thursday under the name K2 Horizon, alongside the weights, training data, code, methodologies, and intermediate checkpoints used to build them — a package the institute's founder frames as a reference point for what "truly open" can mean at frontier scale.
The K2 Horizon family spans from a model sized for smartwatches and other constrained devices up to a 375-billion-parameter system aimed at enterprise deployments. Weights are the tunable numerical values inside a model; intermediate checkpoints are the snapshots a model passes through during training. Publishing both lets outside researchers retrace the development process and reproduce results, not merely fine-tune a finished network.
Founder Eric Xing told Reuters the release is meant to show that openness and competitive performance "are not mutually exclusive." The package goes further than the "open-weight" approach used by some Chinese developers, which makes model files available for download but keeps the underlying data and techniques hidden. It also contrasts with OpenAI and Anthropic, which release neither their models nor the data behind them.
The launch fits the UAE's stated push to position itself as a global AI hub. Whether a single institute's release resets industry practice is an open question; OpenAI and Anthropic have not signaled any change, and the partial-disclosure pattern remains dominant in frontier AI.