IFM Launches K2 Horizon Open-Source AI Model Fleet
The MBZUAI-backed Institute of Foundation Models has launched K2 Horizon, a six-model open-source AI fleet that provides weights, code, and training data to enable true reproducibility.
The Institute of Foundation Models (IFM), an AI research lab backed by MBZUAI, has launched K2 Horizon, a suite of six foundation models ranging from 0.9 billion to 375 billion parameters. Unlike typical open-weights releases, IFM is publishing the entire pipeline under an Apache 2.0 license, including model weights, source code, training data, and recipes. This level of transparency allows developers and researchers to audit training data for licensing issues, verify benchmark claims, and customize models without reverse-engineering hyperparameters.
The fleet features four dense models at 0.9B, 3.7B, 7B, and 32B parameters, alongside two sparse models: a 36B model activating 4B parameters and a 375B flagship activating 23B parameters. IFM claims state-of-the-art performance for its smaller models, positioning the 7B model as the strongest under 10B parameters and the 3.7B model as the leader under 4B parameters on reasoning, math, and coding benchmarks. The 0.9B model targets highly constrained edge devices like smartwatches, while the 3.7B and 7B models are optimized for mobile phones and on-device software engineering.
To improve inference speeds, K2 Horizon introduces a diffusion distillation technique that generates blocks of tokens in parallel, which IFM claims delivers a 3x throughput increase without sacrificing quality. Additionally, the suite debuts a Mixture of Value Attention (MoVA) architecture. Unlike standard Mixture-of-Experts systems that route tokens to different feed-forward networks, MoVA applies sparsity to the value projections of the attention mechanism. This architecture powers the 36B-A4B model, allowing it to run efficiently on workstation-class GPUs while outperforming larger dense models.
For practitioners, the K2 Horizon family is available immediately on Hugging Face, with deployment support via vLLM and SGLang, and API access through AWS, Cerebras, Nebius, and Compass. Because the entire family shares a unified architecture and vocabulary, developers can prototype on the 0.9B model and scale up to the 375B flagship without rewriting deployment code. Furthermore, the Apache 2.0 license contains no commercial revenue thresholds, offering a legally cleaner alternative to other open-weights models for enterprise applications.
This is our own summary of reporting by AlphaSignal
