Mk1.5 explained: factor world knowledge into tool calls, parallel research via sub-agents

AkshatS07 · x · 2026-09-25

The Mk1.5 team outlines two principles behind its real-time model thesis: (1) stronger reasoning lets efficient models solve complex problems by thinking longer, and (2) world knowledge is cleanly factored into external tool calls like multimodal search and reverse image search. A demo shows Mk1.5 spawning parallel sub-agents to research a basket of snacks, with the team extending this to general sub-agent parallelization.

Related event: Mk1.5 Real-Time Multimodal Reasoning Model Launches with 2-5x Lower Latency(4 posts)→

Original post →

More from coding & agent

coding & agent channel →