Anthropic published tests on 12 October on whether small AI models could copy themselves onto cheap devices, such as smart ovens and off-brand Android TV boxes. In a demo, a 26B-parameter model ran across 15 TV boxes, each holding a slice, at 0.39 tokens per second. In a lab network of 8 emulated devices, GLM-5.2 spread to all 8 in four of five trials, and to 7 in the other.
GLM-5.2 is not a small model. It ran on a separate server, because the lab nodes were too small, and it set up Llama3.1-8B on each node. Anthropic uses it as a capability floor, meaning a small model matching it could do the same. A trend in a general capability score, which Anthropic calls only a rough proxy, puts a small model there by the end of 2026. Small replicators may follow soon after. Models up to 30B parameters handle many single steps today but err when joining them. The test made the break-in trivial, so Anthropic calls it an upper bound.
Anthropic calls setting up a model on unfamiliar hardware the least studied step, and possibly the hardest. It says this needs no root access. Buyers and operators can ask which devices have open admin ports or default logins, and who patches them.