Short answer: K2 Horizon is a family of six AI models released on September 3, 2026, by the Institute of Foundation Models at the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) in Abu Dhabi. The models range from 0.9 billion to 375 billion parameters, and the institute describes the release as the largest fully open model release in AI history, meaning it publishes not only the model weights but also the code, training data, and methodology, under the Apache 2.0 license. The models are built for reasoning, mathematics, coding, and agentic tasks, and the smallest is designed to run on a watch. For anyone trying to understand how AI is trained to reason, a release that shows its work is unusually instructive, with one caveat: for the two largest models, the data and training code are promised rather than published yet.
Most AI announcements describe a product. This one describes a process, which is why it deserves attention from people who will never download a model file. When a lab publishes its training data and recipes, the mechanics of how a machine learns to reason stop being a trade secret and become something that can be inspected, reproduced, and argued about in public.
What MBZUAI actually released
The fleet contains six models, each aimed at a different kind of hardware:
| Model | Designed for |
|---|---|
| 0.9B | Highly constrained devices such as watches and glasses |
| 3.7B | Phones and on-device use; described as fine-tuning friendly |
| 7B | Phones; the institute calls it the best-performing model under 10B parameters |
| 32B (dense) | Laptops, local hosting, and on-premise servers |
| 36B, 4 billion active | A new "Mixture of Value Attention" design that activates only 4B parameters per task |
| 375B, 23 billion active | The flagship, built for enterprise reasoning and agentic work |
All six share a core architecture, vocabulary, and tooling, so a developer can prototype on a small model and move to a larger one without rebuilding the workflow. The models are available through Hugging Face, the vLLM and SGLang serving frameworks, and inference partners including Cerebras, AWS, and Nebius.
The flagship's Hugging Face model card adds detail the announcement leaves out: it stores 375 billion parameters but runs only 23 billion for each token, and it accepts a context window of 512,000 tokens. It also exposes a separate reasoning channel, returning its "thinking" apart from its final answer when a developer requests a high reasoning effort.
"Open weights" versus "fully open"
The distinction MBZUAI keeps stressing is easy to miss. Weights are the numbers a model learns during training, and publishing them lets anyone run the model. They say almost nothing about how the model was made. A release that publishes the training data, the code, and the recipe lets outside researchers verify claims, study failure modes, and reproduce results.
Professor Eric Xing, founder of the institute and president of MBZUAI, framed the goal this way: "We believe meaningful AI progress depends on the ability to examine, build upon, and improve the technology, not simply access it through an API."
That aspiration is only partly fulfilled today. The model card for the 375B flagship states that its intermediate checkpoints, data, and training code "will be released," with the technical report and code repository expected by the end of September 2026. The smaller models are further along. Independent reviewers, including CellCog, have also noted that the announcement discloses no figures for the compute used in training. Neither point sinks the release. It does mean separating what you can download today from what is still on the way.
How AI models are trained to reason
Reasoning, in the sense AI labs now use the word, means a model working through intermediate steps before committing to an answer, rather than producing the most statistically likely reply in a single pass. Modern reasoning models are generally trained in stages. A model first learns language and general knowledge from enormous amounts of text. It is then refined on examples of careful problem solving, and it is often further trained with reinforcement learning, in which the model's attempts at problems with checkable answers, such as mathematics and code, are rewarded when they turn out correct. Over many rounds, the model learns habits that produce correct answers more reliably: breaking a problem into parts, checking intermediate results, and abandoning an approach that fails.
K2 Horizon's announcement names three engineering choices that bear on this process:
- Diffusion distillation, which generates blocks of tokens in parallel rather than one at a time, and which the institute says speeds output by roughly three times without degrading quality. Speed matters for reasoning because reasoning consumes tokens; a model that thinks longer before answering costs more time and money per question.
- Mixture of value attention, an architecture the institute says improves reasoning without adding computation.
- Dynamic model routing, which sends each task to the most cost-effective model in the fleet, so a simple request never occupies the flagship.
The third idea is the one with the broadest practical consequence. Most business questions do not require the most powerful model available, and a system that recognizes the difference spends accordingly.
Why openness matters for reasoning research
A reasoning claim is only as credible as the ability to test it. When training data is hidden, a high benchmark score raises an uncomfortable question: did the model reason its way to the answer, or had it effectively seen the answer during training? Published data lets outside researchers check for that kind of contamination. Published intermediate checkpoints, which the flagship's card promises, let them study how reasoning ability emerges across the course of training rather than only at the end.
This matters outside universities too. According to Stanford's 2026 AI Index, as of March 2026 the top closed model led the top open model by 3.3%, up from 0.5% in August 2024. The gap is small enough that choosing between open and closed systems increasingly turns on cost, privacy, and control rather than raw capability.
What this means for businesses that will never train a model
Few small businesses will ever download a 375 billion parameter model, and none need to in order to benefit from this release. The effects reach them anyway, just indirectly.
First, capable reasoning is moving onto small devices. A model that runs on a phone or a laptop can process sensitive information, such as client records, without sending it to an outside service. Medical practices, law firms, and financial advisers have waited a long time for that.
Second, the price of reasoning keeps falling. Faster generation, sparse models that activate a fraction of their parameters, and routing that sends routine work to small models all reduce what each answer costs. Those savings eventually reach the tools businesses already pay for.
Third, the useful question has shifted. Whether AI can reason well enough for everyday business work is largely settled. What remains is plumbing: connecting a capable model to the phone line, the calendar, the CRM, and the follow-up sequence so that the reasoning produces a booked appointment rather than an impressive paragraph. For a broader orientation, see what small business owners need to understand about AI now.
Where Velora fits
Velora Media does not build foundation models. It connects them to the parts of a business that lose money when nobody responds: the missed call, the unanswered inquiry, the lead nobody followed up with. The AI Automation & Integration package starts at $2,000 and deploys an AI receptionist, workflow automation, CRM lead scoring, and review automation as one connected layer, with the published price on the page.
Frequently asked questions
What is K2 Horizon?
K2 Horizon is a family of six AI models from MBZUAI's Institute of Foundation Models, released September 3, 2026. The models range from 0.9 billion to 375 billion parameters and are built for reasoning, math, coding, and agentic tasks under the Apache 2.0 license.
What does "fully open" mean for an AI model?
A fully open model publishes its weights, training code, training data, and methodology, so others can inspect and reproduce it. An "open weights" model publishes only the trained weights. For K2 Horizon's largest models, some data and code are still scheduled for release.
How are AI models trained to reason?
Reasoning models typically learn language from large text collections, then are refined on worked problem solving and often trained with reinforcement learning, which rewards correct answers on checkable problems. Over time the model learns to work through intermediate steps before answering.
Can K2 Horizon run on a phone?
Yes. MBZUAI says its 3.7B and 7B models are designed for phones and on-device applications, and its 0.9B model for more constrained devices such as watches and glasses.
Is K2 Horizon better than ChatGPT?
MBZUAI claims top performance within each size class, and says its flagship is built to compete with leading open-weight models. Stanford's 2026 AI Index found the best closed model led the best open model by 3.3% as of March 2026.
Can a business use K2 Horizon commercially?
The models and code are released under the Apache 2.0 license, which permits commercial use. Running the larger models still requires significant hardware or a paid inference provider.
Related reading
For another view of how quickly AI capability is advancing, see GPT Astra and 3D home modeling. For the hardware side of the AI buildout, see what Micron's $10B AI investment means for small business. For the safety debate running alongside all of this, see Anthropic's AI risk warning.
Velora Media landing pages, by industry
Each site below carries the same six pages: an overview, the services built for that industry, a shop with published prices, the engine showing how the pieces connect, results, and the questions owners in that industry ask first.
Home services: plumbing, electrical, HVAC, and appliance repair
homeservices.veloramedia.cloud · Services · Shop · Engine · Results · FAQ
AI receptionist, for any industry
aireceptionist.veloramedia.cloud · Services · Shop · Engine · Results · FAQ
Real estate agents
realestateagents.veloramedia.cloud · Services · Shop · Engine · Results · FAQ
Real estate wholesalers
wholesale.veloramedia.cloud · Services · Shop · Engine · Results · FAQ
Med spas and aesthetic clinics
medspaaesthetics.veloramedia.cloud · Services · Shop · Engine · Results · FAQ
Physicians and medical practices
physicians.veloramedia.cloud · Services · Shop · Engine · Results · FAQ
Auto body and collision repair
autobody.veloramedia.cloud · Services · Shop · Engine · Results · FAQ
Law firms and attorneys
lawfirms.veloramedia.cloud · Services · Shop · Engine · Results · FAQ
Florists and flower shops
florist.veloramedia.cloud · Services · Shop · Engine · Results · FAQ
Sources
- MBZUAI, "MBZUAI's Institute of Foundation Models launches K2 Horizon, the world's largest fully open AI models in history," September 2026. mbzuai.ac.ae
- Hugging Face, "IFM/K2-Horizon-375B-A23B" model card. huggingface.co
- CellCog, "K2 Horizon: Six Open Models, Licenses, and What Is Missing." cellcog.ai
- Stanford HAI, "The 2026 AI Index Report: Technical Performance." hai.stanford.edu