TUESDAY, NOVEMBER 17, 2026 Tech news, minus the noise.
N News Cast
LATEST
Home  /  AI
AI

K2-Horizon Open Models Are the Most Open Release of 2026

IFM's K2-Horizon fleet ships six open models, from a 0.9B dense checkpoint to a 375B mixture-of-experts, with training data, checkpoints, and code included.

K2-Horizon Open Models Are the Most Open Release of 2026
Source: Ana Las Heras, CC BY-SA 4.0, via Wikimedia Commons

K2-Horizon open models arrive with everything included

The K2-Horizon open models landed between November 15 and 17, and they reset what an open release is supposed to mean. IFM, the Abu Dhabi-based institute behind the K2 line, published six models ranging from a 0.9B dense checkpoint to a 375B-A23B mixture-of-experts. The weights are only the start. The release also includes training data, intermediate checkpoints, and the infrastructure code used to train them, all published at ifm.ai/k2 alongside the paper numbered 2611.21003.

Most open-weight releases hand you a finished model and a model card. This one hands you the factory. IFM says it is releasing the pipelines that produced roughly 10 trillion synthetic pretraining tokens, not just the merged checkpoints. That is a different kind of openness, and it matters most to the people who actually build on top of models rather than just calling an API.

What the fleet actually contains

The fleet covers six sizes, and each one has a job. The dense models run 0.9B, 3.7B, 7.0B, and 32B parameters. The mixture-of-experts pair is a 36B-A4B model built for cost-efficient agentic work and the 375B-A23B flagship aimed at coding and agentic workloads. All four of the larger models share a 22 trillion-token pretraining corpus, which is the kind of scale that used to belong only to closed labs.

The technical contributions are specific, not marketing words. MoVA attention and softplus attention gating change how the model routes information, and the Uno diffusion adapters add a new way to adapt the models after training. The Hugging Face paper page lays out these contributions in detail, which is exactly the point: you can read the paper, download the data, and check the claims yourself.

Why full openness changes the math

Open weights alone let you run a model and fine-tune it. Open data and open training code let you audit it, reproduce it, and improve it in ways the original authors never tried. That distinction is the whole story here. A team that distrusts a benchmark can rerun the training recipe. A researcher who suspects contamination in the data can look at the data. A startup that needs a specialized variant can start from an intermediate checkpoint instead of the finished model.

This is also why the release landed the way it did. K2-Horizon follows K2 Think V2 from January 2026 in IFM's public lineage, and it reads as a generation refresh with heavier openness obligations, not a silent weight bump. The institute is betting that the builders who care about provenance will pick the fleet that shows its work.

It arrives into a crowded month. Alibaba launched Qwen4-Max on November 16 with 2.4 trillion parameters, Anthropic shipped Claude Fable 5.2 a week earlier, and Meta's Muse Spark Open showed that an 816B-parameter model can run on two Mac Studios. K2-Horizon is not trying to win that size race. It is trying to win the trust race.

The catch: size, residency, and cost

None of this is free, and the flagship makes that obvious. A 375B-A23B mixture-of-experts is not a laptop download. Running it takes serious GPU hardware, and training on the released data takes far more. The smaller models are the practical entry point for most teams, and the 36B-A4B MoE is probably the one that gets the most real use.

There is also a geography question. IFM is Abu Dhabi-based, and some teams will need to map data residency and export rules before mirroring full corpora, separate from any question of model quality. That is normal for any release of this scale, but it is worth checking before you build a pipeline around the data.

Who should care, and who should wait

If you build agents, fine-tune models, or research training methods, this release is directly for you. The combination of six sizes, shared pretraining data, and published pipelines is the most complete open package of the year. Start with the 7B or 36B models, read the paper, and test the checkpoints against your own workload before committing.

If you just need a model behind an API, nothing here changes your week. The hosted options from the big labs are still easier. But the next time someone tells you an open model is just a weights dump, point them at K2-Horizon. The bar moved, and it moved a lot.

AS
News Cast Editorial Team

We cover AI, gadgets, startups and security — original reporting and explainers, written by humans, for humans. Got a tip? Reach us via the contact page.