Institute of Foundation Models releases K2 Horizon, six fully open models from 0.9B to 375B parameters

The Institute of Foundation Models (IFM), the model-building arm of Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), released K2 Horizon on September 3, 2026 — a family of six language models, from 0.9 billion to 375 billion parameters, shipped as fully open: weights, code, training data, and training methodology all published together.

What's new

Each K2 Horizon model card frames the release around a specific role in a shared family rather than a single flagship number. As IFM's own Hugging Face page for the mid-size model puts it: "K2-Horizon-7B is the medium dense member of the K2-Horizon family: a 7B-core decoder-only model with a 512K context window." The six sizes span very different deployment targets — small models built for running locally or on a single node, mid-size dense models for cost-sensitive production use, and a Mixture-of-Experts design (K2-Horizon-MoVA-36B-A4B) that stores 36B parameters but activates only 4B per token, plus a 375B model aimed at long-horizon agentic workloads.

All six are released under the Apache 2.0 license and are downloadable directly from Hugging Face under the IFM organization, alongside GGUF quantized builds for local inference via llama.cpp, and support for vLLM, SGLang, and Ollama serving. IFM has also published intermediate training checkpoints for several sizes, letting outside researchers study how capabilities develop over the course of training rather than only inspecting a finished model.

Context

Most "open" model releases from major labs ship weights alone, with training data and methodology kept proprietary. K2 Horizon's pitch is to go further: publishing the training data and recipes alongside the weights, which IFM and outside coverage have framed as one of the largest fully open model launches to date, spanning six coordinated sizes sharing core architecture, vocabulary, and evaluation infrastructure rather than being trained independently.

The release lands into a crowded open-weights field that already includes Meta's Llama line, Alibaba's Qwen family, DeepSeek, and Mistral's open releases — but most of those disclose weights and sometimes technical reports without the full training data and code IFM is publishing here.

Why it matters

For researchers and smaller labs, a fully open release with real training data is worth more than another weights-only checkpoint: it lets outside teams reproduce results, audit what a model was actually trained on, and study checkpoints mid-training rather than treating the final model as a black box. The size spread — from a 0.9B model meant to run on a laptop up to a 375B model aimed at long-horizon agent work — also gives developers a single, architecturally consistent family to build on across very different deployment budgets, instead of stitching together unrelated models from different labs at each size tier.