IFM Releases K2 Horizon: Six Apache 2.0 Models From 0.9B to 375B

Most model launches arrive with one checkpoint and a glossy chart. IFM went bigger. The Institute of Foundation Models, the frontier lab MBZUAI launched in May 2025, released K2 Horizon as a six-model family ranging from 375B-A23B down to 0.9B. Alongside the weights came the pre-training corpus, intermediate checkpoints, training code, configs and fine-grained logs. IFM calls it the largest fully open-source model launch in AI history.

The important bit is that this is not just a pile of files. All six models are on Hugging Face under Apache 2.0, with FP8 and GGUF builds. IFM says they can be used on day one with vLLM, SGLang and Ollama, across NVIDIA, AMD and Cerebras hardware. Hosted APIs are also live through Compass, Cerebras and Nebius via platform.ifm.ai. The pitch is portability: teams can start small and move up the stack without changing their serving setup.

Under the hood, the family shares one architecture, one vocabulary, one training method and one deployment stack. The 0.9B model uses a smaller vocabulary, but the broader point stays the same. IFM says each model was pre-trained on roughly 20 trillion tokens, with nearly 17% of the corpus made up of problem-solving trajectories that include explicit reasoning. About 10 trillion tokens were synthetic, and post-training data was mixed in during mid-training instead of being held back for the end. The team also says it created more than 100 million unique synthesized tasks. Tool definitions were shown in JSON, XML and Markdown during training, with Markdown becoming the default at inference because it was roughly 18.5% more token-efficient than JSON on IFM’s data.

The more interesting technical twist is MoVA, short for Mixture-of-Value Attention. Traditional mixture-of-experts only sparsifies the feed-forward layers. MoVA pushes routing into multi-head attention too, which gives the model another place to scale capacity. The 36B-A4B version has 36B total parameters with about 4B active per token, and under matched training conditions it comes in slightly below the dense 32B model. IFM’s tables still show it leading its comparison set on Terminal-Bench 2.1 and tau3-Banking.

IFM also shipped Uno, a LoRA adapter that freezes the autoregressive model and trains diffusion parameters to generate blocks of tokens in parallel. The company claims a roughly 3x speedup with no quality loss. And then there’s the self-audit, which is the part more labs should copy. IFM ran 375B-A23B on 89 Terminal-Bench 2.1 tasks with eight attempts each, then re-audited the passes using Artificial Analysis’s reward-hacking procedure. The audit flagged 24 trials across 10 tasks, cutting reported accuracy from 70.2% to 66.9%. That’s a cleaner number, and a better habit, than pretending benchmarks are sacred.