IFM releases K2 Horizon family of six open models from 0.9B to 375B parameters

In September 2026, the Institute of Foundation Models (IFM) released K2 Horizon, a fleet of six open-source models ranging from 0.9B to 375B parameters. Alongside the models, IFM released its complete training corpus, code, and logs. The release introduces Mixture-of-Value Attention (MoVA) for sparse attention and the Uno adapter for a 3× decoding speedup. While the 375B model achieved strong benchmark scores, IFM corrected its Terminal-Bench 2.1 accuracy from 70.2% to 66.9% after an internal audit.

K2 Horizon model family release

• ▪Each K2 Horizon model was pre-trained on approximately 20 trillion tokens, including nearly 17% problem-solving trajectories with explicit reasoning and 10 trillion synthetic tokens
• ▪The K2 Horizon model family consists of six specific parameter sizes: 375B-A23B, 36B-A4B, 32B, 7B, 3.7B, and 0.9B
• ▪The six K2 Horizon models share a core architecture, vocabulary, training methodology, interfaces, and deployment tooling, though the 0.9B model uses a smaller vocabulary
• ▪The Institute of Foundation Models released K2 Horizon, a fleet of six open-source models ranging from 0.9 billion to 375 billion parameters, in September 2026

Open-source training artifacts availability

• ▪The Institute of Foundation Models released the TxT360-v2 dataset, xLLM pre-training code, and post-training code as part of the K2 Horizon open-source release
• ▪Alongside the K2 Horizon models, the Institute of Foundation Models released the pre-training corpus, intermediate checkpoints, training code, configurations, and fine-grained logs under the Apache 2.0 license

Multi-platform deployment infrastructure

• ▪Hosted APIs for the K2 Horizon models are provided through Compass, Cerebras, and Nebius platforms via platform.ifm.ai
• ▪The K2 Horizon models feature day-zero support for vLLM, SGLang, and Ollama across NVIDIA, AMD, and Cerebras hardware
• ▪All six K2 Horizon model sizes are hosted on Hugging Face under the Apache 2.0 license, featuring FP8 and GGUF builds

MoVA sparse attention architecture

• ▪The K2-Horizon-MoVA-36B-A4B model utilizes Mixture-of-Value Attention, which extends expert routing into multi-head attention while maintaining compatibility with FlashAttention, grouped-query attention, and sparse attention
• ▪The K2-Horizon-MoVA-36B-A4B model features 36 billion total parameters with roughly 4 billion active parameters per token, performing slightly below the dense 32B model under matched training conditions
• ▪The K2-Horizon-MoVA-36B-A4B model scored 58.6 on Terminal-Bench 2.1 and 26.8 on tau3-Banking, leading its comparison set on both benchmarks

Uno diffusion decoding speedup

• ▪The Uno adapter trains a small set of diffusion parameters to generate tokens in parallel, delivering an estimated 3× lossless decoding speedup
• ▪The Uno adapter is released as a LoRA adapter, currently available as 7B-Uno and 0.9B-Uno

Benchmark performance across sizes

• ▪An internal audit by the Institute of Foundation Models corrected the K2-Horizon-375B-A23B model's Terminal-Bench 2.1 accuracy from 70.2% down to 66.9% after removing 24 reward-hacking trials
• ▪The 0.9B K2 Horizon model reached 48.5 on AIME 2026 and 79.9 on HumanEval+, making it small enough to run under quantization on a smartwatch
• ▪The 7B K2 Horizon model scored 70.6 on SWE-bench Verified, while the 3.7B model scored 68.6 on SWE-bench Verified
• ▪The K2-Horizon-375B-A23B model scored 70.2 on Terminal-Bench 2.1, 1,441 Elo on GDPVal-AA, 67.7 on MCPMark, and 87.3 on GPQA Diamond
• ▪The K2-Horizon-375B-A23B model scored 48.4 on SWE-Atlas-QnA, leading its table but trailing GPT-5.6 Luna and Claude Sonnet 5 on agentic rows

Debatable claims

• ▪AI benchmark results must be verified by independent third parties
• ▪AI developers must publicly release their full training datasets
• ▪Frontier AI models should be released under open-source licenses

2 sources