// AI NEWS · MODEL RELEASE

MBZUAI's Institute of Foundation Models Launches K2 Horizon, Six Fully Open AI Models From 0.9B to 375B

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

Published 2026-09-03 · Updated

Short answer

K2 Horizon is a family of six open AI models from MBZUAI's Institute of Foundation Models, released September 3, 2026. The sizes run from 0.9B to 375B parameters, and IFM published the weights, code, training data and methodology, with models and code under Apache 2.0. They're text-only and built for reasoning, math, coding and agentic tasks. The 375B flagship has a 512K-token context window.

Key takeaways

  • Six sizes: 0.9B, 3.7B, 7B, 32B, 36B-A4B and 375B-A23B.
  • Weights, code, training data and methodology are all public.
  • IFM calls it the largest fully open model release in AI history.
  • Each model was pretrained on roughly 20 trillion tokens.
  • Artificial Analysis currently scores the 375B at 31, against a median of 19 for comparable open models.

What is K2 Horizon?

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 small enough for a phone and has 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 currently lists the 375B at 31 on its Intelligence Index, against a median of 19 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.

Why does a fully open model family matter for creators?

  • Fully open — not just open-weight. Publishing training data, code, and methodology alongside the weights is rarer than an open-weight release, and it lets developers and researchers reproduce, audit, and fine-tune the models rather than just run them.
  • A size for every deployment. Six models sharing one architecture and training recipe means you can prototype on a 7B, serve production on a 36B or 375B, and run something on the edge from the same family — with consistent behavior as you scale up or down.
  • It is a text and reasoning family, not a content tool. K2 Horizon writes, reasons, and codes; it generates no images, video, captions, or scheduled posts. For creators, that draws a clean line between "drafting engine" and "content operation."
  • Apache 2.0 lowers the cost floor. Unrestricted commercial use plus self-hosting means high-volume drafting, extraction, and agentic work can run on your own infrastructure instead of per-token API fees — the same economics that made open models attractive for automation.
  • Model choice keeps commoditizing. As capable open families multiply, the model itself becomes a swappable component, and the durable advantage shifts to the workflow, brand voice, and distribution wrapped around it.

How do you turn a big model launch into content?

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 for X and LinkedIn, a Quote Graphic pulling the "six models, fully open, Apache 2.0" line, a Carousel breaking down what each size is for, a captioned Persona Short 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 and published across the eight social platforms plus blog and email from a single queue on 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.

What are the key facts about K2 Horizon?

  • September 3, 2026: MBZUAI's Institute of Foundation Models released K2 Horizon — six open text models from 0.9B to 375B parameters.
  • Fully open under Apache 2.0: weights, source code, training data, and methodology, with unrestricted commercial use.
  • Sizes target different jobs — 0.9B for the edge up to the 375B-A23B MoE flagship for long-horizon agents, with a 512K-token context on the 375B.
  • Text-only, tuned for reasoning, math, coding, and agentic tool use; each model pretrained on roughly 20 trillion tokens.
  • It is a drafting-and-reasoning engine, not a content tool. Kompozy is what turns the news — or a draft — into captioned, on-brand posts across nine platforms.

What does this look like in Kompozy?

A real quote graphic generated in the Kompozy workspace, shown as an example of the format. It wasn't made about this story. One line from a launch take can carry a post like this.

Quote graphic generated in Kompozy
Real output · Kompozy Quote graphic format · from the Kompozy workspace

Who is this for, and who should skip it?

Best for:

  • Developers who want to self-host and fine-tune open models
  • Researchers who need training data and code, not just weights
  • Teams that want one model family from edge devices to servers

Not ideal for:

  • Creators who need images, video or captions, since the models are text-only
  • Anyone who wants a hosted content tool rather than a model

Frequently asked questions

What is K2 Horizon?

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.

What does "fully open" mean here?

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.

Can K2 Horizon create social media content?

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.

How big is the largest K2 Horizon model?

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.

How do we know this?

We checked every date, figure and quote on this page against the sources listed below on 2026-10-23, including MBZUAI's press release and Artificial Analysis's live model page; the benchmark score was updated to the figure listed there today, and rewrote or removed anything they didn't support. Kompozy plan and feature details come from kompozy.io/pricing.

Sources

  1. MBZUAI press release via Zawya — K2 Horizon launch (read 2026-10-23)
  2. Artificial Analysis — K2 Horizon 375B A23B (read 2026-10-23)
  3. NYU Shanghai RITS — IFM releases K2 Horizon (read 2026-10-23)
  4. Kompozy — Pricing (read 2026-10-23)

Related news

← All AI news · Get started →