Overview
The Abu Dhabi-based Foundation Model Institute (IFM, affiliated with MBZUAI and led by Eric Xing) released K2 Horizon on September 3, 2026, a family of six fully open models (0.9B to 375B-A23B), all under the Apache 2.0 license. This release not only makes model weights public but also unusually opens up training code, data mixing ratios, intermediate checkpoints, and fine-grained training logs, adhering to the full-open principle since LLM360. The flagship 375B-A23B is a sparse MoE architecture with native 512K context support, performing strongly on benchmarks like Terminal-Bench 2.1. The release coincided with NVIDIA's acquisition of Hugging Face, and the Apache 2.0 license ensures model portability and neutrality.
Key Features
- Fully Open and Reproducible: Beyond weights, it publishes training code, data mixing ratios, data construction recipes, intermediate checkpoints, training configurations, and fine-grained logs. For datasets under license restrictions, it provides source descriptions and filtering processes to ensure reproducibility, described by Reuters as one of the most complete open-source releases to date.
- Unified Architecture and Seamless Migration: All six models share the same architecture, vocabulary, training recipe, interfaces, and evaluation tools, allowing teams to seamlessly migrate deployments across sizes, from watch-level devices to enterprise-grade servers.
- Flagship Sparse MoE and Long Context: The 375B-A23B stores 375B parameters, activates about 23B per token, and supports native 512K context (524,288 tokens), targeting enterprise-level long-range reasoning and agentic tasks.
- Proprietary MoVA Sparse Attention: The 36B-A4B and 375B-A23B adopt IFM's self-developed Mixture-of-Value Attention mechanism, pushing sparsity to the attention value layer. The 36B-A4B activates only about 4B parameters per token, improving inference efficiency.
- Large-Scale Synthetic Data Training: Each model is pre-trained on approximately 20 trillion tokens, with about 10 trillion being synthetic data (about 17% explicit reasoning traces), and post-trained on over 100 million generated tasks, enhancing reasoning and tool-use capabilities.
- Day-0 Ecosystem Support: Immediate support on vLLM, SGLang, Ollama, accessible via APIs from Compass, Cerebras, AWS, Nebius, etc. Weights are hosted on Hugging Face, with Uno Diffusion adapters for 2.5-3x inference acceleration.
Use Cases
- Enterprise-level long-range reasoning and agentic workflows, leveraging the 375B-A23B's 512K context for complex multi-step tasks.
- Edge and mobile applications, using 3.7B or 7B models for on-device intelligence on phones or edge devices.
- Local workstations and private server deployments, using 32B models in data-sensitive environments.
- Efficient inference services, using the 36B-A4B sparse model to reduce computational costs while maintaining performance.
- Lightweight AI on constrained devices (e.g., watches), with the 0.9B model meeting extremely low-resource requirements.
Pros
- Fully open license (Apache 2.0), with weights, code, and data recipes freely accessible and mirrorable.
- Flagship model achieves 70.2% on Terminal-Bench 2.1 and 42.6% on SWE-bench Pro, showing strong tool-use and coding capabilities.
- Self-audit mechanism proactively identifies and fixes benchmark issues, enhancing credibility.
- Unified architecture and interfaces simplify deployment migration from edge to cloud.
- MoVA sparse attention mechanism reduces inference costs while maintaining quality.
- Immediate support from mainstream inference frameworks on release day, enabling rapid ecosystem integration.
Pricing
Official pricing information is not publicly disclosed; model weights are free and open-source, and inference API prices are subject to actual quotes from service providers (e.g., Compass, Cerebras, AWS, Nebius).
Summary
K2 Horizon is IFM's fully open model family, covering six sizes from 0.9B to 375B-A23B, under Apache 2.0 license with full training pipeline disclosure. The flagship model excels with sparse MoE and 512K context, performing well on coding and agentic benchmarks. The release emphasizes reproducibility and ecosystem neutrality, suitable for diverse deployments from edge to enterprise.
Version History
- IFM K2 Horizon launched (six fully open models) (2026-09-03): The Institute of Foundation Models (IFM) at MBZUAI in Abu Dhabi released K2 Horizon, a family of six fully open models from 0.9B to 375B-A23B, all under Apache 2.0. Beyond weights, IFM published training code, data recipes, intermediate checkpoints and evaluation logs. The 375B-A23B flagship activates ~23B params per token with a native 512K context and scores 70.2% on Terminal-Bench 2.1. Day-zero support shipped for vLLM, SGLang and Ollama, alongside Uno Diffusion, a lossless LoRA adapter delivering 2.5-3x faster inference.