K2 Horizon: The First Fully Open Model Fleet

On September 3, the Institute of Foundation Models (IFM), part of MBZUAI, released K2 Horizon: six language models ranging from 0.9 billion to 375 billion parameters, all under the Apache 2.0 license. What makes this release different from the usual model drop is not the size range, it is what got published alongside the weights: the training code, the training data itself where redistribution licenses allow it, intermediate checkpoints, training configurations, and evaluation logs.

Most releases marketed as "open" hand you a set of weights and a model card, the approach Meta took with Muse Glimmer earlier this quarter. K2 Horizon hands you the whole recipe.

What is actually in the release

Six model sizes cover a wide range of hardware budgets:

• 0.9B and 3.7B, small enough to run on a laptop or a modest cloud instance

• 7B and 32B, the range most indie builders would actually self-host

• 36B-A4B, a mixture-of-experts model with 4B active parameters per token

• 375B-A23B, the flagship, mixture-of-experts with 23B active parameters

Alongside the weights, IFM published the training code, training data where licensing allows redistribution, training configurations, intermediate checkpoints, and evaluation results. According to IFM's own release announcement, the smaller tiers hit state-of-the-art results at their size class on reasoning, math, coding, and agentic benchmarks. The models are live now on Hugging Face and served through inference partners including Compass, Cerebras, AWS, and Nebius, per HPCwire's coverage.

Why full openness matters more than open weights

Weights alone let you run a model. They do not let you understand why it behaves the way it does, reproduce its training, or safely fine-tune it without wondering whether your derivative work steps on an undisclosed data license. Publishing the training data and code closes that gap. If you are building a product on top of an open model rather than just calling an API, that difference is the entire point.

It also matters for a narrower, practical reason: builders who fine-tune open-weight models today are frequently guessing at what the base model already saw during training, which makes it hard to tell whether a fine-tune improved a capability or simply duplicated data the model already had. Training data transparency removes that guesswork.

What this means if you are not training models

If your job is shipping a product, not researching foundation models, the release still matters at the small end. The 0.9B and 7B tiers are realistic to self-host cheaply, and a state-of-the-art result at that size class means a smaller, cheaper model doing more of what you previously needed a larger hosted API for. That is a real cost lever for anything running high-volume, low-complexity inference: classification, extraction, simple agent tool routing.

The 375B flagship is not something most teams building an app will ever run themselves. It matters to the field more than it matters to your infrastructure bill this quarter.

The honest limitation

Open weights, even fully open ones, do not remove the operational work of running a model: you still need to serve it, monitor it, and handle the failure modes any model has. "Fully open" is a licensing and transparency claim, not a claim that self-hosting becomes effortless. If your current setup calling a hosted API already works and is not a cost problem, this release is worth knowing about, not necessarily worth migrating for today.

For more on reading model announcements like this one without losing a day to hype, see our guide on keeping up with AI news. If terms like mixture-of-experts and active parameters are new to you, our explainer on how AI models work covers the basics.

FAQ

Is K2 Horizon actually free to use commercially?

Yes. Apache 2.0 permits commercial use, modification, and redistribution, which covers the model weights and the released code.

How is this different from Meta's or Mistral's open-weight releases?

Those releases typically publish weights and a model card only. K2 Horizon adds the training code, training configurations, and the training data itself where its license allows redistribution, which lets you inspect and reproduce the training process, not just run the result.

Can I fine-tune K2 Horizon for my own product?

Yes, and the released training data and configs make it easier to understand what the base model already learned before you fine-tune on top of it.

Which size should I self-host?

For most application-level use (classification, extraction, simple agents) the 7B or 32B tiers are the realistic starting point on typical cloud GPU instances. The 375B flagship needs serious multi-GPU infrastructure.