IFM Launches K2 Horizon: A Fully Open, High-Performance AI Model Fleet

IFM has announced the release of K2 Horizon, a suite of six AI models designed to deliver leading performance across a range of sizes and applications. The models, which include sizes from 0.9 billion to 375 billion parameters, excel in areas such as reasoning, mathematics, coding, and agentic tasks. Notably, the 0.9B, 3.7B, and 7B models set new benchmarks in their respective categories. K2 Horizon stands out for its unprecedented openness. IFM is providing full transparency into the entire training lifecycle—from pretraining through advanced reasoning and agentic post-training phases. This includes releasing intermediate checkpoints, training data or detailed data construction recipes, open architecture designs, mixture compositions, training code, configurations, fine-grained logs, evaluation results, and final model weights. The models and code are available under the Apache 2.0 license, while datasets are shared under their respective licenses or with detailed documentation when redistribution is restricted. The fleet covers a broad spectrum of deployment scenarios. The smallest model, 0.9B, targets highly constrained devices like smartwatches and glasses. The 3.7B and 7B models are suitable for phones and other on-device applications. Larger models such as the dense 32B and sparse 36B-A4B offer powerful options for local workstations and efficient serving, with the 36B-A4B model leveraging a novel Mixture-of-Value-Attention (MoVA) mechanism to achieve high efficiency. The largest, 375B-A23B, is designed for demanding enterprise workloads, featuring a sparse mixture-of-experts architecture that activates a subset of its 375 billion parameters per token. All models share a consistent architecture, vocabulary (with a smaller set for the 0.9B model), training methodology, interfaces, evaluation infrastructure, and deployment tooling. This uniformity facilitates dynamic routing of tasks between models and enables detailed study of capability and efficiency across scales. K2 Horizon’s training involved approximately 20 trillion tokens per model, combining diverse sources including web text, code, scientific and mathematical data, multilingual content, and synthetic data generated through advanced pipelines. A significant portion of the training corpus incorporates explicit reasoning trajectories, enhancing the models’ problem-solving abilities. IFM also developed new metrics to quantify data diversity and carefully documented the data construction process. The release includes detailed training logs and intermediate checkpoints, allowing researchers to analyze training dynamics and the emergence of capabilities. This transparency extends to post-training phases, where the models undergo supervised fine-tuning, reinforcement learning, and agentic specialization. The development process produces a branching model tree rather than a single endpoint, enabling examination of how reasoning, tool use, planning, and agentic skills develop. To address inference speed challenges, IFM introduces Uno Diffusion, a lossless speedup technique that accelerates token generation without compromising quality. Uno uses diffusion distillation to generate token blocks in parallel while preserving the autoregressive model’s output distribution, offering improved latency and throughput for interactive and large-scale applications. The entire infrastructure used to build K2 Horizon, including the xLLM training system and agentic post-training code, is also being released. This enables researchers and developers to reproduce training stages, adapt models to new domains, and extend capabilities. IFM emphasizes that K2 Horizon is more than a set of models; it is an open scientific experiment. For example, the team audited the largest model on TerminalBench 2.1 tasks to identify instances of reward hacking—where the model exploits benchmark artifacts rather than genuinely solving problems. Such behaviors can now be studied in detail thanks to the availability of intermediate checkpoints. All six K2 Horizon models are available under open licenses with support for deployment on NVIDIA, AMD, and Cerebras hardware. The release aims to provide a transparent foundation for advancing AI research and development across scales and applications.