When most AI companies say "open source," they mean you can download the weights. The Institute of Foundation Models (IFM) at Abu Dhabi's MBZUAI just raised the bar. On September 3, 2026 (UTC), they launched K2 Horizon — six models from 0.9B to 375B parameters, every single one released with weights, training code, training data, intermediate checkpoints, and fine-grained training logs. All under Apache 2.0.
Six models, one architecture
| Model | Parameters | Active params | Type | Best for |
|---|---|---|---|---|
| K2 Horizon 0.9B | 0.9B | 0.9B | Dense | Watches, glasses, edge devices |
| K2 Horizon 3.7B | 3.7B | 3.7B | Dense | Fine-tuning, on-device |
| K2 Horizon 7B | 7B | 7B | Dense | Phones, laptops |
| K2 Horizon 32B | 32B | 32B | Dense | Local hosting, on-premise |
| K2 Horizon 36B-A4B | 36B | 4B | MoVA | Efficient server inference |
| K2 Horizon 375B-A23B | 375B | 23B | MoE | Enterprise reasoning, agents |
The three smallest models set new state-of-the-art results in their size classes across mathematics, reasoning, software engineering, and agentic tasks. The 0.9B model is compact enough to run on a watch. The 7B model, IFM claims, is the best-performing model under 10B parameters and can run on a phone.
The 375B flagship is a Mixture-of-Experts model with 23B active parameters per token, built to compete with leading open-weight models on reasoning and agentic work. All six models share the same core architecture, vocabulary, and deployment tooling, so developers can prototype on the 0.9B model and scale to the 375B without changing workflows.
What "fully open" actually means here
This isn't open weights with a restrictive license. K2 Horizon ships with:
- Model weights for all six models
- Training code and configurations
- Training data mixtures and recipes (including pretrain and mid-train data on Hugging Face)
- Intermediate checkpoints at various training stages
- Evaluation logs and methodology
IFM calls it the largest fully open model release in AI history. Given that recent "open" releases from major labs have stopped at weights-only with commercial restrictions, that claim is hard to argue with.
Three technical innovations
K2 Horizon isn't just a bigger Llama clone. IFM built in three notable techniques:
- Diffusion distillation — generates blocks of tokens in parallel, speeding inference by roughly 3x without degrading response quality.
- Mixture of Value Attention (MoVA) — a new architecture used in the 36B model that improves reasoning while activating only 4B parameters.
- Dynamic model routing — directs tasks to the most cost-effective model in the fleet, giving developers a practical path from prototype to production.
Why this matters
IFM was founded in May 2025 (UTC) and now operates from Abu Dhabi, Silicon Valley, and Paris. It's led by Professor Eric Xing, MBZUAI's president, and draws talent from the world's leading AI centers. The UAE has been aggressively building its AI credentials — from G42's infrastructure deals to HUMAIN's Arabic models — and K2 Horizon is another signal that Abu Dhabi wants to be taken seriously as a frontier AI research hub, not just a buyer of Nvidia chips.
For developers, the appeal is obvious: a fully transparent model fleet from edge to enterprise, Apache 2.0 licensed, with no API lock-in. The models are available now on Hugging Face, vLLM, and SGLang, with API access through Compass, Cerebras, AWS, and Nebius.
Whether the 375B model can actually compete with Llama 4 or DeepSeek on benchmarks remains to be seen. But IFM has made it easy for anyone to verify — because they released everything needed to reproduce the results.
Sources: MBZUAI, PR Newswire, AI/TLDR, The National, Singularity.Kiwi
No comments yet