The Institute of Foundation Models released a connected fleet of six foundation models including full training code, data recipes and checkpoints. The flagship K2 Horizon 375B uses a Mixture-of-Experts architecture with 23B active parameters and a 512K token context window. It recorded a score of 47 on the Artificial Analysis Intelligence Index, marking a 30 point increase over the previous 70B parameter K2 Think V2.
Smaller versions of the family range from 0.9B to 36B parameters to support local hosting on devices including watches and smartphones. The 375B flagship decreased its hallucination rate to 26% from 71% in the previous generation by declining to answer 60% of questions on the AA-Omniscience benchmark. Day-0 support for the fleet is live in SGLang.
Key sources
- SOURCE@ifm_ai“largest fully open-source model launch in AI history”x.com
- SUPPORT@kimmonismus“training code, data recipes, checkpoints, logs and evals”x.com
- SUPPORT@artificialanlys“score of 47 on the Artificial Analysis Intelligence Index”x.com
- SUPPORT@artificialanlys“reduced hallucination is a huge improvement over K2 Think V2 (71%)”x.com
- SUPPORT@sgl_project“0.9B (runs on a watch), 3.7B, 7B (runs on a phone)”x.com