MBZUAI K2 Horizon unveils six fully open models up to 375B

K2 Horizon is a landmark for the open‑weight ecosystem and a statement of intent from the Gulf. MBZUAI is not just releasing another 7B or 70B model; it is shipping a coordinated family that stretches from sub‑billion‑parameter edge models to a 375B‑parameter MoE system, with Apache‑2.0 licensing and a full training recipe. That combination of scale, openness and institutional backing moves the center of gravity for serious open research beyond the usual US‑EU axis and gives labs worldwide a credible alternative to heavily redacted frontier models. ([mbzuai.ac.ae](https://mbzuai.ac.ae/news/mbzuais-institute-of-foundation-models-launches-k2-horizon-the-worlds-largest-fully-open-ai-models-in-history/?utm_source=openai))

For the race to AGI, K2 Horizon matters because it lowers the barrier to experimenting with truly frontier‑scale architectures outside big tech. When researchers can inspect weights, data curation and RL regimes on a 375B‑class model, they can test new alignment schemes, interpretability tools and agentic training loops without waiting on closed labs. At the same time, pushing fully open models into this capability range will sharpen policy debates around export controls and dual‑use risk, especially given the model’s strong coding and agentic benchmarks. Expect K2 Horizon to become a reference point in future arguments over whether open source helps or hurts safety, and a catalyst for other state‑backed institutions to mount their own open frontier runs. ([wam.ae](https://www.wam.ae/en/article/c227c2u-mbzuais-institute-foundation-models-launches?utm_source=openai))