IFM Announces K2 Horizon, Deploying Six Models with Full Training Process Disclosure

IFM stated that these models achieve top-tier performance for each of their respective sizes. Specifically, the company announced that the 0.9B, 3.7B, and 7B models recorded state-of-the-art (SOTA) performance for their scale in the domains of mathematics, inference, coding, and agent tasks.

In addition to the model release, a key feature is the disclosure of the entire training process. Intermediate checkpoints, training data and data composition recipes, open architecture, mixture configurations, training code, detailed logs, evaluation results, and the final weights are all being made public.

The models are provided under the Apache 2.0 license. Datasets follow applicable licenses, such as ODC-BY. In cases where redistribution is not possible, the methods for data construction and mixing will be disclosed.

The largest model, 375B-A23B, employs a sparse MoE (Mixture-of-Experts) architecture. While the total parameter count is 375 billion, approximately 23 billion parameters are activated per token.

IFM stated that this model ranks among the top for models with fewer than 400 billion parameters in terms of general capabilities, inference, coding, and agent evaluations. It is designed for complex inference, software engineering, and long-term tasks.