Released September 3, 2026, K2 Horizon is a family of six text models spanning 0.9B to 375B parameters, all published with weights, code, training data, and methodology under Apache 2.0 — a claim to the largest fully open AI release to date.
2026-09-03 · by Moe Ameen
The Institute of Foundation Models (IFM) at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) released K2 Horizon on September 3, 2026 — a family of six open large language models ranging from 0.9 billion to 375 billion parameters. What sets the launch apart is not just scale but the degree of openness: IFM published each model's weights, source code, training data, and methodology under the Apache 2.0 license, which permits unrestricted commercial use. IFM, with teams in Abu Dhabi, Paris, and Silicon Valley, is positioning K2 Horizon as the largest fully open AI model release to date.
The six sizes are pitched at deliberately different jobs. The 0.9B is built for edge devices — think wearables and on-device use — with IFM claiming best-in-class math, reasoning, and tool use for its size class. The 3.7B targets mobile and rapid fine-tuning, the 7B is aimed at smartphones and developer workstations with a software-engineering bent, and the 32B is a dense model for high-memory laptops and on-premises servers. The two largest use sparse mixture-of-experts designs: the 36B-A4B activates about 4B parameters per token and adds a mixture-of-value attention mechanism, while the flagship 375B-A23B activates roughly 23B per token and is engineered for long-horizon agentic work and complex reasoning.
These are text-only models — no image, audio, or video generation — tuned for reasoning, mathematics, coding, agentic tool use, and long context. IFM says each model was pretrained on roughly 20 trillion tokens, and the 375B ships with a native 524,288-token (512K) context window. On third-party evaluation, Artificial Analysis rates the 375B at 47 on its Intelligence Index against a median of 29 for comparable open-weight models. Treat individual specs and benchmark figures as a launch-window snapshot and confirm them against IFM's own model cards and repositories before relying on any single number.
A model launch this big is a story your audience is already talking about, and for a creator the fastest win is not running K2 Horizon — it is being early and clear on what it means. That is a distribution opportunity, and it is exactly the kind of raw material [Kompozy](/) turns into a week of on-brand content. Drop the angle you want to take — "the largest fully open AI model just shipped, here's why open training data matters" — into Kompozy and it fans that one take into a native [Text Post](/glossary/content-repurposing) for X and LinkedIn, a [Quote Graphic](/glossary/hyperframes) pulling the "six models, fully open, Apache 2.0" line, a Carousel breaking down what each size is for, a captioned [Persona Short](/glossary/persona-shorts) explaining it to camera, a Blog Article for the long tail, and a newsletter to your list.
Every piece is written in your voice through one [Persona Brief](/glossary/persona-brief) and published across the eight social platforms plus blog and email from a single queue on [Autopilot](/glossary/autopilot), with a per-post review gate before anything ships. The point of the pairing is timing: K2 Horizon is the news; Kompozy is how you own the conversation around it while it is still fresh — the model reasons over text, and the content engine turns your commentary into finished posts everywhere. For context on how open-weight models are spreading into mainstream tooling, see [GitHub Copilot adding Kimi K2.7 Code](/news/kimi-k2-7-github-copilot).
K2 Horizon is a family of six open large language models released on September 3, 2026 by the Institute of Foundation Models (IFM) at MBZUAI. The models span 0.9B to 375B parameters and are published fully open — weights, code, training data, and methodology — under the Apache 2.0 license. They are text-only, tuned for reasoning, math, coding, and agentic tool use.
IFM released not only the model weights but also the source code, the training data, and the training methodology under Apache 2.0. That goes further than a typical "open-weight" release, which usually ships weights alone — it lets researchers reproduce, inspect, and adapt the models, and permits unrestricted commercial use and self-hosting.
No. K2 Horizon is a text and reasoning model family — it writes, reasons, and codes, but generates no images, video, captions, or scheduled posts. To turn a draft or a news take into on-brand content published across platforms, you pair it with a content engine like Kompozy, which generates across formats and publishes to nine social platforms plus blog and email.
The flagship is the 375B-A23B: a sparse mixture-of-experts model with about 375 billion total parameters and roughly 23 billion active per token, shipping with a native 524,288-token (512K) context window. IFM positions it for long-horizon agentic work and complex reasoning.