MBZUAI publie K2 Horizon, six modèles fully open jusqu'à 375 B
TL;DR
- Six models from 0.9B to 375B share a common architecture and ship simultaneously under Apache 2.0 with weights, data, training code, and methodology — the most complete public disclosure on record at this parameter scale.
- The 7B tier posts a 20-point SWE-bench Verified gain over prior open-weight SOTA at that scale, making it the practical breakout for single-GPU and on-device agent deployments, per CellCog.
- IFM self-published a reward hacking correction, dropping a benchmark score from 70.2% to 66.9% after third-party audit — Moor Insights identifies this voluntary disclosure as the launch's strongest credibility signal, not its biggest number.
L'Institute of Foundation Models de MBZUAI a mis en ligne jeudi K2 Horizon, une famille de six modèles ouverts allant de 0,9 à 375 milliards de paramètres, diffusés avec les poids, le code, les données d'entraînement et la méthodologie complète. L'université d'Abou Dhabi le présente comme "one of the largest 'fully open' models in AI history", selon The National — formulation plus prudente que la revendication de première place que l'on entendait autour du lancement.
Hector Liu, directeur du laboratoire de l'IFM en Silicon Valley, indique que le plus petit modèle est "small enough to run on a watch", tandis que le modèle phare à 375 B est "built to compete with leading open-weight models on reasoning and agentic work". "Every model is built to compete with the best open models at its size, and everyone ships with the weights, code, training data and methodology behind it," ajoute-t-il. L'article ne publie aucun benchmark.
Le président de MBZUAI Eric Xing inscrit le geste dans une position de principe : "We believe meaningful AI progress depends on the ability to examine, build upon and improve the technology, not simply access it through an API." La suite embarque un "dynamic model routing" censé orienter chaque tâche vers "the most cost-effective model", et une "mixture of value architecture" annoncée comme un gain de raisonnement sans surcoût de calcul. Les six modèles sont disponibles sur Hugging Face, Ollama, Unsloth et le dépôt IFM.
C'est la deuxième annonce d'un modèle frontière issu du Golfe que nous suivons cette semaine, après Humain-m3.
Ce qu'en disent les autres médias
-
Moor Insights & Strategy Lire →
Tier-1 analyst framing centers on voluntary reward-hacking disclosure and third-party auditing as the credibility differentiator, not raw benchmark scores.
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.
-
ZAWYA Lire →
Gulf regional wire adds UAE sovereign-AI context: MBZUAI is training 80,000 federal employees as AI experts, positioning the state as originator rather than adopter.
The most important technology of our time should be built with the world, not kept from it.
-
CellCog Lire →
Practitioner analysis argues the 7B model is the genuine breakthrough for on-device and single-GPU use, and independently documents the reward hacking self-audit that most coverage ignored.
The 7B is the real headline for anyone running agents on-device or on a single GPU, the small end of the fleet matters more.
Article original publié par thenationalnews.com
Lire l'article original →Titre original : MBZUAI publie K2 Horizon, six modèles totalement ouverts de 0,9 à 375 B de paramètres