CONFIRMED: the Institute of Foundation Models (IFM) at MBZUAI (Mohamed bin Zayed University of Artificial Intelligence, in Abu Dhabi) has released K2 Horizon, a family of six open language models ranging from a 0.9-billion-parameter dense model to a 375-billion-parameter mixture-of-experts (MoE) model. The announcement was published in November 2026, with a 134-page technical report on arXiv on 16 November and coverage on 17 November.
The release stands out not only for its size: alongside the weights, IFM published the intermediate training checkpoints, the pre-training data, the run records and the training and evaluation code, following the LLM360 definition of open science. That is 380 checkpoints, 13 TB of pre-training data, pipelines that generated roughly 10 trillion synthetic tokens, records from 58 training runs, 17 code repositories and the Search360 index, with 994 million web documents used to ground synthetic-data generation.
Six models, from pocket to data center
The fleet has four dense models — 0.9B, 3.7B, 7B and 32B — and two mixture-of-experts models: the 36B-A4B, with about 4.6 billion active parameters per token, and the flagship 375B-A23B, with 375 billion parameters in total and about 23 billion active per token. Four of the six models (3.7B, 7B, 32B and 36B-A4B) share a single 22-trillion-token pre-training corpus and the same training recipe, enabling controlled comparisons of scale and sparsity. The 375B model was trained separately, on 15.1 trillion tokens in two phases, on NVIDIA H200 GPUs, and ships in FP8 and NVFP4 quantizations; the smaller models also come in GGUF and MLX builds for local inference.
Benchmarks: leading on tool use, with caveats
According to figures reported by the authors themselves — with no independent replication yet — the 375B-A23B scored 65.3 on Toolathlon-Verified, a multi-step tool-use test, ahead of GLM-5.2, MiniMax M3, Nemotron 3 Ultra and Inkling, and above GPT-5.6 Terra. The model also scored 48.4 on SWE-Atlas-QnA, a repository-level coding test. The authors themselves acknowledge the main gap: expert-level scientific reasoning, on the HLE and CritPt tests. And one correction is worth noting: the Terminal-Bench score was revised down from 70.2 to 66.9 after auditing.
MoVA, softplus gating and Uno decoding
On the architecture side, the paper introduces Mixture-of-Value Attention (MoVA), which applies mixture-of-experts routing to attention value projections, combined with softplus attention gating. With both, the 36B-A4B nearly matches the 32B dense model trained on the same data while using about one-seventh of the active parameters. The team also shipped Uno, a diffusion decoding adapter that speeds up generation by up to 2.2 times without a separate draft model, available for the 0.9B and 7B models.
STATEMENT: “Science works when others can see the data, follow the method, reproduce the result, and improve on it,” said Eric Xing, founder of IFM, presenting the release. The artifacts are published on the project's official page and in the K2 Horizon collection on Hugging Face, under the Apache 2.0 license for models and datasets.
One important caveat: the two largest models, 375B and 36B, currently ship with weights only; their full training data and code were promised but have not been published yet. Anyone deploying the models in production should check each artifact's license and hardware requirements before rolling it out.
Also from MBZUAI: on 11 November, Geekiki reported on PAN 2, a 29B open-world model that simulates physical actions in video.
Image credit: MBZUAI (Mohamed bin Zayed University of Artificial Intelligence) logo, via Wikimedia Commons (CC BY-SA 4.0).
Sources
- IFM — K2 Horizon: Open-Source AI Models for Every Scale (página oficial)
- AI Weekly — MBZUAI Opens K2 Horizon Fleet, Six Models From 0.9B to 375B
- AlphaSignal — IFM Opens K2 Horizon’s Full Training Stack Across Six AI Models
- Daily AI Brief — MBZUAI’s Institute of Foundation Models releases K2 Horizon, six fully open models up to 375B parameters
This text was produced with the assistance of artificial intelligence, cross-checking public posts and internet sources. Consult the original sources to confirm the information.
Por Muse