MBZUAI published six K2 Horizon models with weights, data and training methods
The models run from 0.9 billion to 375 billion parameters under Apache 2.0, with vLLM, SGLang, Ollama and Unsloth support from launch. Publishing the training data is the rare part.
MBZUAI's Institute of Foundation Models has released six K2 Horizon models ranging from 0.9 billion to 375 billion parameters, under Apache 2.0, publishing weights, code, training data and training methods. They are supported by vLLM, SGLang, Ollama and Unsloth from launch, according to The National.
Most releases described as open publish weights and nothing else. Publishing the training data and the methods is a different standard, and very few organisations meet it.
Why the data release is the substantive part
A model with published weights can be run, fine-tuned and inspected. It cannot be understood, reproduced or audited, because the thing that determines its behaviour — what it was trained on — is undisclosed.
Publishing the corpus makes several things possible that are otherwise not. Researchers can investigate why a model behaves as it does by looking at what it learned from. Benchmark contamination can be checked directly rather than inferred. Bias can be traced to sources instead of characterised only in outputs. And another group can reproduce the training run, which is the basic requirement for a scientific claim.
It also creates legal exposure, which is why almost nobody does it. A published corpus is a published list of what you used — inspectable by every rights holder in it, at a moment when the US Justice Department has filed in support of OpenAI's fair use position, the Seattle Times and Newsday have sued over training data, and authors are contesting the division of Anthropic's copyright settlement.
An institution funded by the UAE government has a different risk calculus from a company facing US class actions.
The size range is a deployment decision
Six models from 0.9 billion to 375 billion parameters covers the practical spectrum rather than competing at a single point.
The 0.9 billion model runs on a phone. The 375 billion model needs a serious cluster. Between them sit the sizes most production systems actually want, and releasing the family together lets a team develop against one and deploy on another without changing the tokeniser, the prompt format or the fine-tuning recipe.
Launch support for vLLM, SGLang, Ollama and Unsloth matters more than it sounds. A model that requires custom serving code is a research artefact; one that works with the standard inference stack on day one is deployable that afternoon. The gap between those two has killed more open releases than quality has.
Where this leaves the open-weight map
The serious open-weight publishing this year has come almost entirely from Chinese labs and Gulf-funded institutes.
Alibaba refreshed Qwen3.8-Max at 2.4 trillion parameters. Z.ai released GLM-5.3-Flash, 320 billion parameters with 18 billion active, at a tenth of its predecessor's cost. OpenBMB released MiniCPM5-2B under Apache 2.0. And MBZUAI has now published a full family with its training data.
Meanwhile the American and European labs have moved in the other direction, towards gated capability tiers for vetted customers: Anthropic's Mythos 5.1, OpenAI's Daybreak cohort, Google's Fairwind programme.
That is a coherent safety position and it has a consequence. The models that researchers can actually study — weights, data, methods — are increasingly the ones published by institutions outside the jurisdictions writing the safety rules.
MBZUAI has not published the compute used or independent evaluations of the models.
Runs the newsroom. Rename this profile in the studio to your own byline.
Related
Every weekday, the AI stories that moved money or shipped code.
No cross-posting, unsubscribe anytime. See all newsletters