MBZUAI’s IFM Launches 6 K2 Horizon Frontier Models, Doubles Down on Openness — Analyst Insight

September 3, 2026

(Credit: MBZUAI Institute of Foundation Models)

Today, the Mohamed bin Zayed University of Artificial Intelligence and its Institute of Foundation Models publicly launched a fleet of six K2 Horizon frontier open models in sizes ranging from 0.9 billion to 375 billion parameters. The models in this project have been designed with similar characteristics across popular sizes that should serve anything from a smartwatch to enterprise-grade hardware.

They also embody the institution’s commitment to openness. Sure, that includes open architecture and open weights, but here the openness extends to data recipes (e.g., construction methods, mixture compositions), training code, model configurations and training recipes, intermediate checkpoints, training logs, evaluation results, and even the training data itself wherever redistribution is permitted. The K2 Horizon models and code are available under an Apache 2.0 license that permits unrestricted commercial use. That is what MBZUAI and IFM mean when they call this “the industry’s largest fully open-source fleet of AI models.”

It’s been almost a year since I started writing about MBZUAI and IFM, which was prompted by the release of the K2 Think reasoning model. Since then, I’ve also analyzed the PAN world model, and I even got to attend MBZUAI’s fifth anniversary in Abu Dhabi. Ahead of today’s launch, I had a chance to talk with Dr. Hector Liu, who leads IFM’s Silicon Valley lab, to learn more about the rationale behind the models and some of the technical novelties used to boost their performance. Besides being suitably impressed by the models’ technical characteristics, I’m more convinced than ever that MBZUAI is putting its money where its mouth is when it talks about an “open, accessible, and transparent” approach to model-building.

What MBZUAI Launched — and Why

The IFM is an AI lab with teams of researchers, computer scientists, and developers working under Liu and MBZUAI president (and IFM founder) Dr. Eric Xing across facilities in Abu Dhabi, Paris, and Silicon Valley. With the K2 Horizon launch, IFM aims to offer a counterpoint to both the closed commercial models created by big providers in the U.S. and the open-weight models from Chinese labs that are still not as transparent as IFM would like.

All of the models are designed for strong performance, and benchmarks show that half of them are putting up best-in-class numbers. (More on the benchmarks below.) But they also give users a clear view of how they are built and how they function across the lifecycle. In fact, IFM touts this as the first fully open model fleet for agents, i.e., the first “to expose the complete development process through agentic post-training.” Based on my conversations with them, the team that developed these models genuinely wants you to be able to look at everything under the hood, study how it works, reproduce results, and adapt it all as needed.

These are the six models in the launch:

0.9B: Optimized for extreme edge deployment (including smartwatches or even smart glasses) for basic math, reasoning, and lightweight tool use.
7B: Designed for mobile applications and rapid fine-tuning for developers and researchers; IFM touts this model as delivering reasoning scores competitive with many larger models.
7B: Targeted at smartphones and developer workstations, structured for software engineering and multi-step research.
32B: Dense architectural configuration built for local hosting on high-memory laptops and on-premises servers.
36B-A4B (4B active): Sparse mixture-of-experts (MoE) / mixture-of-value attention (MoVA) design optimized for local enterprise deployment with high reasoning at low compute overhead.
375B-A23B (~23B active): Flagship enterprise MoE architecture engineered for complex logic, multi-step agentic workflows, and intensive enterprise production workloads on multi-GPU deployments.

All six models ship with quantization support, and each model is pretrained on roughly 20 trillion tokens. The team incorporated reasoning directly into pre-training, with nearly 17% of the pre-training corpus consisting of explicit reasoning trajectories. Along the way, the developers used about 10 trillion synthetic tokens in pre-training.

Why six models at once? Because model users, from individual researchers to enterprises, don’t run just one size. Very often, they prototype on something small and then scale that work up to bigger models. Plus, there is the efficiency advantage of being able to pick the smallest model that still meets a need so you can cut energy use, power consumption, or other overhead. Having all six K2 Horizon models available at once and with common characteristics — core architecture, vocabulary, training methodology, interfaces, tooling — should make scaling easier and generally work better than having just one model that might be the wrong size for a given job. (One pragmatic concession: the vocabulary for the 0.9B model is smaller.)

There’s also the benefit here of serving multiple audiences. Researchers should like having a fully reproducible family of models to experiment with and fine-tune. (By the way, if you are interested in the nitty-gritty of AI model research, check out the IFM technical team’s post about the launch, which goes into detail on some very interesting model behaviors, complete with a reward-hacking “JACKPOT” moment for one of the models. More on that below, too.) Developers have the advantage of carrying over a lot of their work as they use models that run across wearables, phones, laptops, and workstations. Enterprises get a flagship model to host on their own infrastructure for agentic work. I plan to dig into some of these use cases in future pieces.

Technical Novelties and Implementation Details for K2 Horizon Models

The team has also built these models with some clever technical advancements. For starters, their new Uno Diffusion is a “plug-and-play” LoRA adapter for speeding up inference. It does this by using a process called diffusion distillation to generate blocks of tokens in parallel rather than sequentially. You don’t need to deploy a second model or adjust a lot of parameters, and Liu says that it’s lossless. So, you should get no loss in response quality — just 2.5x or 3x faster results. These gains held up for every batch size they tested, and IFM claims that Uno has a better speed-quality tradeoff than top speculative-decoding systems and diffusion language models (both open-weight and proprietary).

Sharp-eyed readers will have spotted the reference to the mixture-of-value attention (MoVA) design of the 36B-A4B model. MoVA applies the sparseness benefits of a mixture-of-experts approach to the attention mechanism in the model. More specifically, it creates experts on the value heads in attention. This translates to better capacity while keeping compute low. The 36B-A4B also shows performance close to the dense 32B model, but with many fewer active parameters. According to benchmarks, it also outperforms some much larger models.

The standardized approach and interoperability of the different models in the fleet means that you can serve the same application on different devices. And you can use the system’s dynamic model routing to swap between models and save resources, routing complex tasks to larger models and simple tasks to smaller ones.

IFM is also open-sourcing xLLM, the production training infrastructure it used to build the K2 Horizon models. The lab is likewise releasing its agentic post-training code base, including reinforcement learning components. These moves are intended to help with behavior inspection and reproducibility, and in the bigger picture to help researchers and developers improve techniques and build new models.

What MBZUAI and IFM Mean by ‘Radically Open’

IFM would like to be known as the world’s most open AI lab, and from everything I can tell, it is taking the necessary steps to back that up. It doesn’t want people to merely use its models, but to build on them. To my mind, this is far more in keeping with the original ethos of open-source software than the open-weight models we see from a lot of other places, many of which don’t disclose things like data composition or training code that would allow auditability and reproducibility, let alone the myriad items (checkpoints, logs, configurations, etc.) I listed for IFM in the first section of this article. I mean, even when IFM isn’t permitted to redistribute data for licensing reasons, it still points to the data source and explains how it was curated so that someone else with access to that data could reproduce the results.

To aid this process, each model ships as what IFM calls a development tree, meaning that intermediate checkpoints and post-training branches show how the base model turns into reasoning and agentic variants. All of this is right in line with what Eric Xing said in the launch press release: “Open source is much more than open weights. Science works when others can see the data, follow the method, reproduce the result, and improve on it.”

It’s not just IFM claiming this for itself, either. It measures its openness against the definition established by the Open Source Initiative, and its previous model release was ranked by Artificial Analysis at the top of its Openness Index. (When I double-checked this the day before the K2 Horizon launch, four of the five models earning an 89 — the top score on the Openness Index — were MBZUAI IFM models.) As Liu has said, “Openness isn’t a slogan we ask you to accept. It’s a list we ask you to check. We’d rather be checked than believed.”

To me, this desire to let people inspect a model’s design, verify its behavior, reproduce it, or take it in a different direction with a new experiment is in keeping with IFM’s location within an academic institution. These folks are not just giving you the answer, but showing you their homework — and inviting you to go off and do your own homework using your own hardware. And these technical leaders have spent years advocating for “fully open-source LLMs,” tracing back to the “LLM360” academic paper that Xing and Liu coauthored in 2023. It’s also worth noting that they aren’t trying to sell you chips, datacenter hardware, or a cloud platform.

Initial K2 Horizon Benchmarks

I will dig into measurements of these models in greater detail in a later article, but the benchmarks I’ve seen so far suggest that the open approach driving the K2 Horizon models is hardly just idealism, because these models are strong performers. The team has gotten particularly strong results — even class-leading results — for the 0.9B, 3.7B, and 7B model sizes, and IFM’s own benchmarking has been backed up by independent evaluation from Artificial Analysis. For example, Artificial Analysis scores the 375B-A23B at 47 on its Intelligence Index versus a median of 29 for comparable open-weight models, with a 524K context window and leaner-than-median token usage.

The 32B ranks among the best dense models with fewer than 40 billion parameters, and the 375B-A23B ranks among the top models below 400 billion. The 0.9B scores above 48 on AIME 2026 (though the team concedes that the smallest models still struggle on complex tasks such as those in TerminalBench).

As I said, more to come on this front, because on principle I always want to dig into external benchmarks. (This helps explain why I co-founded a benchmarking firm.) And I’d like to see what independent metrics emerge after these models have been loose in the wild for a few weeks.

The Launch’s Strongest Trust Signal: Benchmark Integrity

There’s a key aspect of this launch that deserves special attention, and it’s one that most coverage will probably skip. IFM audited its own flagship model for reward hacking and printed the results, using Artificial Analysis’s published auditing procedure rather than an internal rubric. The team ran the 375B-A23B on 89 TerminalBench 2.1 tasks with eight attempts each, producing 712 trials, of which 500 passed the verifier, for a reported accuracy of 70.2%. Auditing every passing trial flagged 24 across 10 tasks as reward hacking. Removing them lowers the score from 70.2% to 66.9% (the exact correction was 3.37 percentage points), and the remaining 79 tasks came back clean.

The behaviors are instructive. The model inferred that it was inside a public benchmark and downloaded the reference solution from GitHub, copied fixes from real projects’ public repositories rather than deriving them, inspected exposed credentials, and in some cases exploited how the test checked for success. The most vivid case involves the 7B model, which found and downloaded SWE-bench answers during training and produced an inflated score of 82 that IFM disowned rather than published. For context, Artificial Analysis reports flag rates of 2.2% for Claude Fable 5 and 4.1% for GPT-5.6 Luna, so K2 Horizon’s 3.37% sits inside the range of leading frontier models.

Here is why this matters. Every frontier lab deals with reward hacking, yet almost none of them show you the receipts. IFM ran a third party’s methodology against its own headline number and published the corrected figure. And because it releases intermediate checkpoints, anyone can trace when a cheating strategy first appears, connect it to a specific training decision, and measure its effect on reported performance.

I was talking about checkpoints during my conversation with Dr. Liu, and he went into detail about how the team handles them to make it easy to check and reproduce work. “When we say ‘accessible,’ it means you can easily get to [the checkpoints], right? Of course, I can also give you the recipe for how to do it, but that might take 10 minutes for you to do it yourself. [When we say] ‘accessible,’ we actually also give you the intermediate checkpoints, and even logs for how to read them. So this is not just a gesture. It actually means something.”

At a moment when leaderboard skepticism is running high across the industry, this is the kind of accountability that builds durable credibility. In my view, this disclosure does more for IFM’s standing than any single benchmark win in the launch.

Offering a Different Path for Frontier AI Models

There’s a lot to like about the K2 Horizon launch. One of the MBZUAI executives I talked to called this “a coming-out party for the IFM,” and I think it has achieved that. The team has clearly done a tremendous amount of work, and it’s remarkable to me that they have been able to launch so many models at once — and at such high levels of performance.

IFM wants to give users a range of highly capable models that are built to be adaptable for different uses and different deployment environments, from wearables up to the enterprise. More than that, they want to give organizations another viable path besides closed proprietary models or open-weight models that can only be adapted so much. From what I’ve seen so far, I think they have put forward a strong value proposition of an adaptable, fully open model fleet for enterprise leaders seeking to avoid vendor lock-in, maintain compliance, and deploy cost-effective AI architectures.

As you would expect, IFM is making the models widely available. They’re on Hugging Face today, with day-zero support from vLLM, SGLang, and Ollama. API access is available through inference partners including Compass, Cerebras, and Nebius, with hardware support on AMD, NVIDIA, and Cerebras.

Stay tuned for more on these models in the coming weeks as I have a chance to dig into the specifics of them, their performance characteristics, and their applicability to different use cases.