Four Mac Studios run trillion-parameter AI model Kimi K2.6

At Apple WWDC 2026, LM Studio teamed up with Apple to run the trillion-parameter Kimi K2.6 model on a cluster of four Mac Studios. The cluster has about 1.5TB of unified memory, using Apple's memory sharing tech to handle the big model's needs.

Kimi K2.6 has 1 trillion total parameters with a MoE architecture, 32 billion active parameters, and supports long context, multimodal input, and agent tasks. In the demo, users could securely access the model remotely from a MacBook Neo or iPhone via LM Link, keeping data and communication local.

Tests show Kimi K2.6 can generate about 28 tokens per second on similar setups, with lower power use than traditional GPU clusters. LM Studio's local AI platform shows it's possible to run large open-source models on consumer hardware, highlighting Apple's progress in on-device AI.