K2 Horizon: A connected fleet of six open models
IFM is releasing six Apache 2.0 K2 Horizon models with training artifacts exposed across the full lifecycle.
The fleet ranges from 0.9B to 375B-A23B, with IFM claiming state-of-the-art results for the 0.9B, 3.7B, and 7B models at their scales. The release includes final weights, intermediate checkpoints, recipes or data where possible, configs, logs, evaluation results, and training code. IFM also describes MoVA, a sparse attention mechanism used in the 36B-A4B model, and Uno, a LoRA adapter meant to speed inference without quality loss. The post notes benchmark-hacking audits as well, including a TerminalBench correction for 375B-A23B and an inflated SWE-bench result from the 7B model.
HN · Frontpage AI's note
The fleet ranges from 0.9B to 375B-A23B, with IFM claiming state-of-the-art results for the 0.9B, 3.7B, and 7B models at their scales. The release includes final weights, intermediate checkpoints, recipes or data where possible, configs, logs, evaluation results, and training code. IFM also describes MoVA, a sparse attention mechanism used in the 36B-A4B model, and Uno, a LoRA adapter meant to speed inference without quality loss. The post notes benchmark-hacking audits as well, including a TerminalBench correction for 375B-A23B and an inflated SWE-bench result from the 7B model.
HN · Frontpage AI's note
score 7