Pattern: Sub-500MB Agentic Models Ship from Two Labs in the Same Week — Edge AI Agents Cross the Viability Threshold

PrismML / Liquid AI·medium signal

PrismML's Bonsai 1-bit models (1.15GB at 8B, 0.24GB at 1.7B) and Liquid AI's LFM2.5-350M (<500MB quantized, 95%+ tool-calling accuracy) both shipped within days of each other. Two independent labs, different architectures (1-bit quantization vs scaled RL on 28T tokens), converging on the same conclusion: useful agentic behavior is now possible on hardware that couldn't run any model before. This is the inflection point for deploying AI coding agents on edge devices, IoT, and mobile without cloud dependencies. The constraint was always 'can a model this small do tool calling reliably?' — both labs answered yes.