K2 Horizon is a family of six fully open models from the Institute of Foundation Models (IFM), ranging from 0.9B to 375B parameters, released alongside training data, recipes, code, intermediate checkpoints and training logs. Recent discussion covers community reports that the 7B and 3.7B variants perform well at small sizes, and a company running K2-Horizon 7B among other models on a local 4x RTX 6000 Blackwell MaxQ server for 40 people. IFM researchers are also holding an AMA on October 5.
Company Serves 40 People with Local AI on 4x 6000 Pro GPUsA company built a local AI server with 4x RTX 6000 Blackwell MaxQ GPUs, 384GB RAM, and an AMD EPYC CPU to serve 40 people, costing about €55k. Models include DeepSeek V4 Flash, Qwen 3.8 27B, and K2-Horizon 7B.
IFM team holds AMA on fully open K2 Horizon modelsResearchers at the Institute of Foundation Models (IFM) will hold an AMA on Oct. 5 about K2 Horizon, a fleet of six fully open models ranging from 0.9B to 375B parameters. Alongside the weights they open-sourced training data, recipes, code, intermediate checkpoints and training logs.
Reddit users praise open-source k2 horizon models' performanceA Reddit thread highlights the open-source k2 horizon models, with users claiming the 7B (and 3.7B) variants outperform muse glimmer at smaller sizes and noting the project's fully open release. The community asks whether others have tested them and suggests they could be a milestone if validated.
The A.I. Daily