Anthropic runs 26B-parameter model across 15 Android TV boxes

Anthropic has published research results on running AI models on low-spec IoT devices and on carrying out a replication procedure in a controlled environment, according to byline.network. The company said the work did not demonstrate a fully autonomous, self-replicating AI worm.

In one experiment, Anthropic connected 15 low-cost Android TV boxes and ran the roughly 26-billion-parameter model Gemma4-26B-A4B in a distributed setup, reaching throughput of 0.39 tokens per second.

A separate experiment tested whether a 9-billion-parameter small model could set up its own AI runtime environment: it succeeded on 8 of 9 hardware types.

In a controlled network of 8 Android virtual devices, the large model GLM-5.2 extended the replication task to all 8 devices in 4 of 5 trials. The model actually installed on those devices was Llama3.1-8B, while GLM-5.2 carrying out the replication ran on a separate server. The devices were also placed in an environment that required no authentication for device-to-device access.

Anthropic said small-model performance is improving but still limits consecutive replication. It projected that, on the existing trend, small models could reach the level needed for self-replication by the end of this year.

Botnet context

In January 2026, security firm Synthient reported that the Kimwolf botnet had infected more than 2 million Android-based devices, mainly TV streaming devices.