K2 Horizon
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 · 2026-10-23 · by Moe Ameen
Published 2026-09-03 · Updated
Short answer
K2 Horizon is a family of six fully open text models from MBZUAI's Institute of Foundation Models, released September 3, 2026. The models run from 0.9B to 375B parameters and ship with weights, code, and training data under Apache 2.0, so commercial use and self-hosting are allowed. The 375B flagship has a 512K-token context and scores 47 on the Artificial Analysis Intelligence Index.
Key takeaways
- Six sizes: 0.9B, 3.7B, 7B, 32B, 36B-A4B, and 375B-A23B.
- Weights, code, and training data are all published under Apache 2.0.
- Each model was pretrained on roughly 20 trillion tokens.
- The 375B has a native 512K context and scores 47 on Artificial Analysis.
- Text only: no images, audio, or video.
What is K2 Horizon?
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.
The six sizes share one architecture and training recipe, which is the practical draw: you can move between them without relearning the model. IFM targets the 0.9B at watches and glasses, the 3.7B and 7B at phones, the 32B and 36B-A4B at local workstations, and the 375B-A23B at enterprise serving.
The two largest are sparse mixture-of-experts models: the 36B-A4B and the flagship 375B-A23B, which is aimed at enterprise reasoning work.
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 512K-token context window. On third-party evaluation, Artificial Analysis rates the 375B at 47 on its Intelligence Index. Because the family is new, 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.
What can you make with K2 Horizon?
- First drafts at volume — blog outlines, post copy, scripts, threads, and email bodies you refine and produce elsewhere
- Summaries and extractions from long source material, using the 375B's 512K context to hold a whole document or transcript in view
- Repurposing-ready text: turn one long piece into per-platform angles, hooks, and captions as raw material
- Agentic and tool-using workflows — research, classification, and multi-step reasoning you self-host on the edge or a server
- A self-hosted, Apache-2.0 deployment of the model itself for cost-sensitive, high-frequency text work
- Code and automations that connect your data and pipelines
How does Kompozy turn K2 Horizon drafts into finished content?
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.
Take a draft K2 Horizon produced, and Kompozy fans it into a real package — a captioned Persona Short where your HeyGen avatar narrates the hook, a brand-exact Carousel 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 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 from a long recording, and infographics. Then 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.
- Run K2 Horizon (self-hosted or via a provider) to bulk-draft the week: a rough blog, a batch of hooks, script skeletons, and extractions from your source material.
- Bring a draft into Kompozy as a source and pick your formats — Persona Short, Carousel, Quote Graphics, per-platform Text Posts, a Blog Article, and a Newsletter.
- Let Kompozy rewrite each format to your voice via the Persona Brief so the machine-drafted batch reads as one consistent creator.
- Add the media K2 Horizon can't make — avatar video, clipped shorts, infographics — from the same source in Kompozy.
- Schedule and publish the whole set across the eight social platforms plus blog and email from one queue on Autopilot, reviewing each post before it ships.
What does this look like in Kompozy?
A real Quote Graphic generated in the Kompozy workspace. One strong line from a draft, turned into a post. The line here was not drafted with K2 Horizon.
Who is this for, and who should skip it?
Best for:
- Teams that want a fully open model they can inspect, fine-tune, and self-host
- Researchers who need the training data, not just the weights
- High-volume text drafting where per-token fees hurt
Not ideal for:
- Anyone without GPUs or a hosting provider, especially for the 375B
- Creators who need images or video from the model itself
Frequently asked questions
What is K2 Horizon?
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.
Is K2 Horizon free to 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.
Can K2 Horizon make videos or images?
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.
How do I turn K2 Horizon drafts into finished content?
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.
Which K2 Horizon model should a creator use?
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.
How do we know this?
We checked every figure on this page against the sources listed below on 2026-10-23 and removed or corrected any claim we could not confirm. IFM's own blog and MBZUAI's news page blocked our fetch, so model sizes, license, training tokens, context, and the Artificial Analysis score were checked against the two launch reports listed below. Kompozy plan and format details come from kompozy.io and kompozy.io/pricing.
Sources
- NYU Shanghai RITS — IFM releases K2 Horizon: six open models with data, code, and logs (read 2026-10-23)
- Pasquale Pillitteri — IFM releases K2 Horizon, six open models from 0.9B to 375B with the training data (read 2026-10-23)
- Kompozy — Home (formats and platforms) (read 2026-10-23)
Related tools
- Kimi K2.7 Code — Moonshot AI's open-weight coding model — now the first open-weight option in the GitHub Copilot model picker.
- DeepSeek V4 — The latest generation of DeepSeek's open-weight model family — a two-tier lineup (V4-Pro and V4-Flash) of mixture-of-experts language models with a 1M-token context, MIT-licensed weights, and API pricing well below Western frontier models.
- Claude Sonnet 5 — Anthropic's cheaper, more agentic mid-tier Claude model — close to Opus 4.8 performance at a fraction of the price.