Moonshot AI released Kimi K2.6, a 1-trillion-parameter open-source MoE model that activates only 8 of 384 expert networks per token for inference efficiency. SiliconAngle reports it tops GPT-5.4 and Claude Opus 4.6 on HLE-Full, a 2,500-question doctorate-level benchmark. The model ships with a 400M-parameter vision encoder, support for up to 300 parallel sub-agents for long-running coding workflows, and weights published on Hugging Face under a Modified MIT License.