A K2-Horizon technical report was submitted to arXiv on November 16. It describes a family of six language models and the development process behind them. This is a new research publication; the report’s date should not be treated as the first release date of every individual model.

The authors describe a range from a 0.9-billion-parameter dense model to a mixture-of-experts configuration with 375 billion total and 23 billion active parameters. The paper is publicly available under CC BY 4.0. We verified access to the report, but have not independently audited the entire claimed collection of weights and training artifacts.

Transparency beyond the finished checkpoint

The described release includes intermediate checkpoints, data, logs, and training and evaluation code. The report discusses infrastructure including xLLM and RL360. The value of such a collection goes beyond running the final assistant: it can help researchers ask how behavior emerged at different development stages.

The paper introduces Mixture-of-Value Attention, applying an expert mechanism to attention value projections, and discusses softplus attention gating. Comparative results are reported by the authors. We have not reproduced training or verified the claimed decoding speedup from Uno adapters.

Why the history can be useful

Our editorial assessment is that a preserved development history provides a better foundation for causal questions. If two systems differ in both data and size, it is difficult to attribute the result to either change. Controlled comparisons between successive versions can reduce some of that uncertainty.

A useful example would examine a particular capability before and after post-training. If answer formatting changes but correctness does not, an overall metric might improve without the model learning the solution. That is a proposed analytical question, not a finding about K2-Horizon.

A complete artifact collection also creates practical requirements. Researchers need to know which version corresponds to which evaluation, which scripts are necessary and what resources they require. An available link does not establish that another team can reproduce the whole process in its own environment.

Reproducibility can therefore be assessed in stages. One team might verify a small evaluation slice; another might examine data processing or a checkpoint transition. Clearly reporting those boundaries is more useful than describing a partial check as a full replication. This is an editorial proposal for using the report, not a claim of completed external validation.

The new publication matters as documentation for that work. The next step is independent use of specific parts of the collection, with a recorded scope. In that sense, an open model can offer a traceable path to its results rather than only a file containing the final weights.