K2 Horizon Ships Six Open Models With Training Data

An Open Model Release That Includes the Training Data

The Institute of Foundation Models released K2 Horizon on September 3, 2026, and the interesting part is not the parameter count. It is the manifest. Alongside six sets of model weights, IFM published the training code, the model configurations, the intermediate checkpoints, the training logs, the evaluation results, and the training data itself wherever redistribution licences allowed it. That is a materially different artefact from what the industry has spent two years calling open source.

• Six models: 375B-A23B and 36B-A4B sparse mixtures of experts, a 32B dense model, plus 7B, 3.7B and 0.9B
• Licence: Apache 2.0 on weights and code
• Training scale: roughly 20 trillion tokens, with about 17% of the corpus made up of reasoning trajectories
• Availability: Hugging Face at launch, with day-zero support in vLLM, SGLang and Ollama, and API access through Compass, Cerebras, AWS and Nebius

What Ships Inside the K2 Horizon Fleet

The six models are sized for distinct deployment targets rather than as a marketing ladder. The 0.9B is aimed at genuinely small hardware — watches and glasses. The 3.7B is positioned for mobile and for teams that want a cheap base to fine-tune. The 7B targets phones and developer workstations with a software-engineering lean. The 32B is a dense model for local hosting on a high-memory box, the kind of thing that fits the hardware our readers keep building in our mini computer coverage.

Above those sit the two sparse models. The 36B-A4B activates about 4 billion parameters per token, and the flagship 375B-A23B activates roughly 23 billion. Both use a mixture-of-experts design, which is why the flagship can be positioned as enterprise-scale without demanding enterprise-scale inference for every request.

IFM itself launched in May 2025 under the Mohamed bin Zayed University of Artificial Intelligence, and now runs labs in Abu Dhabi, Silicon Valley and Paris.

Why Does "Fully Open" Mean More Than Open Weights?

Almost every model marketed as open in 2026 is open-weight. You get the parameters and a licence, and the data composition, the curation decisions and the training recipe stay inside the building. That is useful for deployment and close to useless for reproduction.

Analyst Patrick Moorhead of Moor Insights and Strategy, reviewing the release independently, catalogued what IFM added on top of weights: data recipes, training code, model configurations and training recipes, intermediate checkpoints, training logs, evaluation results, and the training data where redistribution is permitted. Where redistribution is not permitted, IFM names the source and explains the curation method so the corpus can be rebuilt.

IFM founder Eric Xing framed the reasoning plainly: open source is much more than open weights, and science works when others can see the data, follow the method and reproduce the result.

One detail is worth more than the licence text. IFM audited its flagship model using third-party evaluation procedures, found reward hacking, and published corrected figures. Labs almost never disclose that. A release that shows its own failed run is making a much stronger claim about reproducibility than a release that only shows a leaderboard.

How Does K2 Horizon Perform Against Other Open Models?

The headline comparison comes from Artificial Analysis rather than from IFM. The 375B-A23B scores 47 on the Artificial Analysis Intelligence Index, against a median of 29 for comparable open-weight competitors. The 32B ranks among the strongest dense models under 40 billion parameters, and the 0.9B scores above 48 on AIME 2026 — a number that would have been implausible for a sub-billion-parameter model a year ago.

IFM's own claim is that the 0.9B, 3.7B and 7B set state of the art at their respective scales. That is the vendor's account of its own models, and it deserves the usual caution until independent leaderboards catch up. Moorhead's review flags the honest limit: the smallest models still struggle on harder agentic work such as TerminalBench, and external metrics will only firm up after wider deployment.

Two architectural notes explain some of the efficiency. IFM describes a diffusion distillation technique that accelerates generation roughly threefold without quality loss, and a Mixture of Value Attention design intended to make reasoning cheaper per token.

Where Can You Run K2 Horizon Today?

The weights are on Hugging Face now, with day-zero support in vLLM, SGLang and Ollama — which means the 32B and smaller models are runnable on a well-specified workstation the day they landed. Hosted API access runs through Compass, Cerebras, AWS and Nebius, and the models have hardware support across AMD, NVIDIA and Cerebras.

That day-zero local story is becoming the pattern rather than the exception. It is the same shape as Tencent's 770B Apache 2.0 release earlier this week and Kimi K3's 2.8T open weights in July: the weights and the runtimes now arrive together.

Why This Matters for Reproducible AI Research

There is a scoring system for this now. Four of the five models holding the top Openness Index score of 89 are IFM models, which tells you both that the bar exists and that very few labs are clearing it.

The practical consequence is for anyone who needs to explain a model's behaviour rather than just use it. A safety researcher can trace an odd output back to a data source. A university group can rerun the training at a smaller scale and check the recipe holds. A company under audit can say where the corpus came from. None of that is possible with weights alone, and all of it is possible here. That is the release's real contribution, and it is a bigger one than any single benchmark number in our AI coverage this week.

Sources: Institute of Foundation Models — K2 Horizon press release — September 3, 2026; Moor Insights & Strategy — analyst insight — September 3, 2026; AIwire — September 3, 2026.