MBZUAI Ships K2 Horizon: Six Fully Open Models From 0.9B to 375B
MBZUAI's Institute of Foundation Models has released K2 Horizon, a fleet of six Apache 2.0 models ranging from 0.9B to 375B parameters. The lab calls it the largest fully open AI release in history: weights, code, training data, and methodology are all public. The 3.7B and 7B models claim state-of-the-art results at their sizes and are small enough for phones.
On this page
A fleet, not a single model
The Institute of Foundation Models at MBZUAI released K2 Horizon on September 3, and the pitch is scope rather than one flagship. "Rather than release a single model, we're releasing an entire fleet at once," said Hector Liu, director of IFM's Silicon Valley research lab. The six models share a core architecture, vocabulary, training methodology, and deployment tooling, so a team can prototype on the smallest model and scale up without switching ecosystems. The 0.9B model uses a smaller vocabulary than its siblings. [1]
The fleet spans six sizes: 0.9B for highly constrained devices such as watches and glasses, 3.7B for on-device use, 7B sized for phones, a 32B dense model for laptops and on-premise servers, a 36B sparse model that activates only 4B parameters, and a 375B sparse flagship with 23B active for demanding deployments. [1]
Openness beyond the weights
The release leans on a distinction the open-source community cares about: open weights versus fully open. K2 Horizon ships under Apache 2.0 with weights, code, training data, and methodology all published. IFM founder and MBZUAI president Eric Xing framed the philosophy directly: "We believe meaningful AI progress depends on the ability to examine, build upon, and improve the technology," adding that "the most important technology of our time should be built with the world, not kept from it." [1] [2]
The lineage matters too. This is the K2 line that produced K2-Think and K2 Think V2, and K2 Horizon moves past open weights toward end-to-end transparency, which lets outside researchers audit not just what the models do but how they were made. [1]
What the small models claim
For local AI users, the interesting claims sit at the bottom of the range. MBZUAI says the 0.9B, 3.7B, and 7B models set new state of the art at their respective scales: the 0.9B leads its size class in mathematics, reasoning, and tool use, the 3.7B is described as the best reasoner under 4B parameters, and the 7B is claimed as best under 10B with strong software engineering ability. The 32B dense model is pitched as among the most capable dense models for local hosting, and the 36B sparse model outperforms many larger models while activating only 4B. [1]
Two technical tricks make the fleet faster and cheaper to run. Diffusion distillation generates parallel token blocks for roughly a 3x speedup, and the sparse models use a new Mixture of Value Attention architecture to keep active parameters low. [1]
Where to get it, and what to check
All six models are on Hugging Face under IFM's profile, with vLLM and SGLang support at launch, plus API access through inference partners including Cerebras, AWS, and Nebius. [1] [3]
The claims deserve the same scrutiny any benchmark-heavy release gets, and that is rather the point: Liu's pitch is that "developers can prototype on the smallest model, scale to the flagship, and verify every claim we make along the way." With training data and methodology public, independent verification is actually possible this time. For anyone choosing a phone-class model, a fully documented 3.7B and 7B pair with state-of-the-art claims is worth benchmarking against whatever you run today, and the Apache 2.0 license means you can fine-tune and ship the results without asking anyone. [1] [4]
Sources
- MBZUAI's Institute of Foundation Models launches K2 HorizonMBZUAI
- Institute of Foundation Models launches the industry's largest fully open source fleet of AI modelsPR Newswire
- MBZUAI IFM launches 6 K2 Horizon frontier models, doubles down on opennessMoor Insights and Strategy
- MBZUAI's IFM ships K2 Horizon: six fully open models from 0.9B to 375B parametersAI Weekly