Ifm unveils k2 horizon: a suite of six open-source AI models ranging from 0.9B to 3,750B parameters

The Institute of Foundation Models (IFM), a research lab launched by MBZUAI in May 2025, has released K2 Horizon, a series of six open-source models under the Apache 2.0 license. The models range from 0.9B to 375B parameters and are available on Hugging Face with FP8 and GGUF builds. They support deployment on NVIDIA, AMD, and Cerebras hardware through platforms like vLLM, SGLang, and Ollama. All models share a common architecture, vocabulary, and training methodology, enabling seamless scaling from the smallest 0.9B model to the largest 375B-A23B variant. The models were pre-trained on approximately 20 trillion tokens, with 17% of the data consisting of problem-solving trajectories. Post-training data was integrated mid-training, and over 100 million synthetic tasks were used. IFM’s K2-Horizon-MoVA-36B-A4B model achieved 58.6 on Terminal-Bench 2.1 and 26.8 on tau3-Banking, outperforming its competitors. The 375B-A23B model scored 70.2 on Terminal-Bench 2.1 and 87.3 on GPQA Diamond, though it lagged behind GPT-5.6 Luna and Claude Sonnet 5 in agentic tasks. IFM also disclosed audit results for the 375B-A23B model, revealing a 70.2% accuracy rate on 712 trials, with 500 passing. After removing flagged trials, the accuracy dropped to 66.9%, aligning with industry benchmarks. The 7B model achieved 70.6 on SWE-bench Verified, while the 0.9B model reached 48.5 on AIME 2026 and 79.9 on HumanEval+.