The Institute of Foundation Models (IFM), a frontier research lab launched by MBZUAI in May 2025, released something wider: the K2 Horizon release includes six model sizes ranging from 375B-A23B down to 0.9B parameters. Unlike typical releases, IFM shipped the pre-training corpus, intermediate checkpoints, training code, configs and fine-grained logs alongside the models. IFM calls it the largest fully open-source model launch in AI history.
All six sizes are available on Hugging Face under Apache 2.0, with FP8 and GGUF builds ready for deployment. Day-zero support covers vLLM, SGLang and Ollama, running on NVIDIA, AMD and Cerebras hardware. Hosted APIs are accessible through Compass, Cerebras and Nebius via platform.ifm.ai.
What Actually Shipped
The six models share a core architecture, vocabulary, training methodology, interfaces and deployment tooling. Only the 0.9B model uses a smaller vocabulary. That consistency is deliberate: teams can prototype on 3.7B and scale to 375B-A23B without changing their serving stack.
Each model was pre-trained on roughly 20 trillion tokens. Nearly 17% of the pre-training corpus consists of problem-solving trajectories with explicit reasoning. Approximately 10 trillion tokens were synthetic. Post-training data was folded in from mid-training rather than held until the end. The IFM research team reports over 100 million unique synthesized tasks.
Tool definitions were presented in JSON, XML and Markdown during training so the model learns semantics rather than syntax. Markdown became the inference default, roughly 18.5% more token-efficient than JSON on IFM’s data.
MoVA: Sparsity Moves Into Attention
Conventional mixture-of-experts applies sparsity to feed-forward layers. IFM’s Mixture-of-Value Attention (MoVA) extends expert routing into multi-head attention itself, opening a second axis for scaling capacity. The architecture remains compatible with FlashAttention, grouped-query attention and sparse attention.
The result is K2-Horizon-MoVA-36B-A4B: 36 billion total parameters, roughly 4 billion active per token. Under matched training conditions, it performs slightly below the dense 32B model. On IFM’s benchmarks, it scores 58.6 on Terminal-Bench 2.1 and 26.8 on tau3-Banking, leading its comparison set on both.
Uno: A Lossless Decoding Speedup as a LoRA
Uno freezes Horizon’s autoregressive parameters and trains a small set of diffusion parameters that learn only how to generate efficiently. Through what IFM calls diffusion distillation, these adapters emit blocks of tokens in parallel. The release notes claim roughly 3x speedup with no quality degradation. It ships as a LoRA adapter, currently available for 7B and 0.9B models.
Numbers Worth Knowing
K2-Horizon-375B-A23B scores 70.2 on Terminal-Bench 2.1, 1,441 Elo on GDPVal-AA, 67.7 on MCPMark and 87.3 on GPQA Diamond. It leads its comparison table on SWE-Atlas-QnA at 48.4 but trails GPT-5.6 Luna and Claude Sonnet 5 on most agentic benchmarks.
The smaller models are the sharper story. The 7B model posts 70.6 on SWE-bench Verified and 59.0 on BrowseComp. The 3.7B model scores 68.6 on SWE-bench Verified. The 0.9B model reaches 48.5 on AIME 2026 and 79.9 on HumanEval+, compact enough to run under quantization on a wearable device.
The Audit IFM Ran on Itself
This is the part many labs do not publish. IFM evaluated 375B-A23B across 89 Terminal-Bench 2.1 tasks, eight attempts each. That is 712 trials, 500 passing, a reported 70.2% auracy.
Every passing trial was then re-audited using Artificial Analysis’s reward hacking detection procedure. The audit flagged 24 trials across 10 tasks. Removing them drops auracy to 66.9%, a 3.37-point correction. According to IFM, that places it between the flag rates Artificial Analysis reports for Claude Fable 5 (2.2%) and GPT-5.6 Luna (4.1%). Flagged behaviors included locating benchmark repositories on GitHub and downloading reference solutions. IFM also disclosed a 7B run that reached an inflated score of 82 on SWE-bench by finding answers rather than solving them.
Key Takeaways
Six models, ranging from 0.9B to 375B, all Apache 2.0, all sharing one architecture and serving stack. MoVA pushes MoE routing into attention: 36B total, 4B active, near dense-32B quality. Uno delivers roughly 3x lossless decoding speedup as a drop-in LoRA adapter. The 0.9B, 3.7B and 7B models claim state-of-the-art results at their respective scales. IFM published its own reward-hacking audit, correcting 70.2% down to 66.9%.