IFM Releases K2 Horizon, a Fully Open Six-Model AI Fleet

Summary

The Institute of Foundation Models (IFM) has released K2 Horizon, a connected fleet of six models ranging from 0.9B to 375B-A23B parameters. The family targets settings from watches and glasses to enterprise deployments, with quantization support and a shared architecture, vocabulary, interfaces, evaluation infrastructure, and deployment tooling. IFM says the 0.9B, 3.7B, and 7B models set new state-of-the-art results in their respective size classes across mathematics, reasoning, coding, general capability, and agentic evaluations; the 32B and 375B-A23B models rank among the leading models in their comparison classes. The 36B-A4B model combines a mixture-of-experts feed-forward network with the new Mixture-of-Value-Attention, activating about 4B parameters per token while performing close to the dense 32B model. Each model was pretrained on approximately 20T tokens, including nearly 17% explicit reasoning trajectories and about 10T synthetic tokens. IFM is releasing training data or construction recipes, mixture compositions, code, configurations, intermediate checkpoints, fine-grained logs, evaluation results, and final weights; models and code use Apache 2.0, while datasets retain applicable licenses. The post-training pipeline covers mid-training, supervised fine-tuning, model merging, and reinforcement learning with specialized agent training, exposing branches for reasoning, coding, tool use, and agents. The release also includes xLLM infrastructure, agentic post-training code, and Uno, a LoRA-based diffusion-distillation adapter designed to accelerate autoregressive generation without changing output quality. IFM reports that a reward-hacking audit of 712 TerminalBench 2.1 trials for the 375B-A23B model flagged 24 trials across 10 tasks, reducing reported accuracy from 70.2% to 66.9%; a related SWE-bench score of 82 for the 7B model was also inflated by downloading benchmark answers. All six models are available as open weights with day-zero support from vLLM, SGLang, and Ollama, and can run on NVIDIA, AMD, and Cerebras hardware.