The Institute of Foundation Models (IFM) at MBZUAI has released K2 Horizon, a fleet of six fully open-source frontier AI models ranging from 0.9 billion to 375 billion parameters, complete with weights, training code, checkpoints and the underlying training data under an Apache 2.0 license.
The largest fully open-source model launch to date
K2 Horizon spans a 0.9B model designed for watches and glasses, a 3.7B model tuned for scientific coding and fine-tuning, a 7B that IFM says is the best-performing model under 10 billion parameters, a 32B dense flagship for local deployment, a sparse 36B-A4B mixture-of-experts and a 375B enterprise-grade model built to compete with leading open-weight rivals on coding and agentic work. IFM claims the 0.9B, 3.7B and 7B set new state of the art at their respective scales.
A glass box, not a black box
Unlike open-weight releases from OpenAI, Anthropic or Google, IFM is publishing the full pre-training corpus (TxT360-v2 on Hugging Face) as well as its pre-training (xllm) and post-training (horizon-post-train) code on GitHub. That radical transparency lets researchers reproduce, retrain and modify each model, which IFM positions as necessary for verifiable AI.
Available today through Hugging Face, vLLM, SGLang, Cerebras and AWS
K2 Horizon models are available immediately through Hugging Face, vLLM, SGLang and Ollama. The IFM API is live via inference partners including Compass, Cerebras, AWS and Nebius. The launch marks the fourth generation in IFM's K2 timeline, following Amber (Dec 2023), K2 V2 (Dec 2025) and K2 Think (Jan 2026).
The release lands the same week OpenAI began rolling out GPT-6 Astra and Google DeepMind shipped Gemini Robotics ER 2, sharpening the split between closed frontier labs and fully open alternatives coming out of the Gulf.
Reporting based on coverage from AIwire, Moor Insights & Strategy, IFM's own press release and PR Newswire.