K2 Horizon Open-Source AI Model Family Generates Buzz for High Performance

A LocalLLaMA community discussion highlights the release of the K2 Horizon model family, specifically drawing attention to its 7B and 3.7B variants. Developed as a radically open-source release, K2 Horizon provides full access to model weights, training code, data, and agentic post-training pipelines. If real-world performance aligns with initial benchmark claims, K2 Horizon's 7B model could outperform significantly larger models like Meta's 30B Muse Glimmer while remaining light enough for local edge execution. Furthermore, releasing every stage of the training pipeline sets a new standard for transparency in open-weights AI development. Developed by MBZUAI and IFM, the K2 Horizon suite includes six open variants supporting context windows of up to 524k tokens designed for agentic tasks. However, community members remain cautious, seeking real-world testing to confirm the models are not simply over-optimized for benchmark scores.