MBZUAI's Institute of Foundation Models: a family of six fully open text models from 0.9B to 375B, released with weights, code, and training data under Apache 2.0.
Last verified · 2026-09-03 · by Moe Ameen
K2 Horizon is a family of six open large language models from the Institute of Foundation Models (IFM), the model lab at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI). Released on September 3, 2026, it spans 0.9 billion to 375 billion parameters, and every model is published fully open — weights, source code, training data, and methodology — under the Apache 2.0 license, which permits unrestricted commercial use and self-hosting. IFM has teams in Abu Dhabi, Paris, and Silicon Valley and frames the release as the largest fully open AI model family to date.
The six sizes share one architecture and training recipe, which is the practical draw: you can move between them without relearning the model. The 0.9B is built for edge and on-device use; the 3.7B for mobile and rapid fine-tuning; the 7B for smartphones and developer workstations with a software-engineering lean; and the 32B is a dense model for high-memory laptops and on-premises servers. The two largest are sparse mixture-of-experts models — the 36B-A4B (about 4B active per token, with a mixture-of-value attention mechanism) and the flagship 375B-A23B (about 23B active per token), built for long-horizon agentic work and complex reasoning.
These are text-only models. K2 Horizon writes, reasons, codes, and uses tools; it does not generate images, audio, or video. 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. Because the family launched today, treat individual specs and benchmark figures as a launch-window snapshot and verify them against IFM's own model cards before relying on any single number.
Think of K2 Horizon as the cheapest place to make raw text at scale, and [Kompozy](/) as the place that scale becomes finished, on-brand content. Because it is Apache 2.0 and self-hostable, a creator or small team can run K2 Horizon to bulk-draft the week's ideas — a rough blog, twenty hook variations, a batch of script skeletons, extractions from a long transcript — for infrastructure cost instead of per-token fees. But a stack of raw drafts is not a content week; it is input. That is exactly where Kompozy picks up, and the division of labor is clean: K2 Horizon reasons over and produces text; Kompozy generates the media the model can't and ships everything.
Concretely: take a draft K2 Horizon produced and Kompozy fans it into a real package — a captioned [Persona Short](/glossary/persona-shorts) where your HeyGen avatar narrates the hook, a brand-exact [Carousel](/glossary/hyperframes) breaking down the argument, Quote Graphics pulling the sharpest lines, native Text Posts tuned per network, a Blog Article for SEO, and an Email Newsletter — all rewritten to one voice through your [Persona Brief](/glossary/persona-brief) so a batch of machine drafts reads like one creator. Kompozy also produces net-new formats K2 Horizon has no path to at all: avatar and persona video, [clipped shorts](/glossary/clipped-short) from a long recording, and infographics. Then [Autopilot](/glossary/autopilot) schedules and publishes the set across the eight social platforms plus blog and email from one queue, behind a per-post review. K2 Horizon drafts cheap and fast; Kompozy turns the drafts into media and puts them everywhere.
K2 Horizon is a family of six open large language models released September 3, 2026 by the Institute of Foundation Models (IFM) at MBZUAI. It spans 0.9B to 375B parameters and is published fully open — weights, code, training data, and methodology — under the Apache 2.0 license. The models are text-only, tuned for reasoning, math, coding, and agentic tool use.
Yes, under the Apache 2.0 license you can download, run, fine-tune, and use the models commercially at no license cost. The real expense is infrastructure — self-hosting the larger models needs real GPUs, and the flagship 375B is a datacenter-scale deployment rather than a laptop one.
No. K2 Horizon is text-only — it writes, reasons, codes, and uses tools, but generates no images, video, or audio. To turn its text into captioned video, brand-consistent images, carousels, and scheduled posts, you pair it with a content engine like Kompozy.
Use K2 Horizon to draft the raw text cheaply, then bring a draft into Kompozy. Kompozy rewrites it to your brand voice, generates the media the model can't (avatar video, carousels, quote cards, infographics), and publishes the whole set across nine platforms plus blog and email from one queue.
For drafting and repurposing text, the mid-size models (7B, 32B, or 36B) are the practical picks — capable enough for quality copy while cheap to run. The 375B flagship, with its 512K context, is best for long-document reasoning and agentic work rather than routine drafting.