Anthropic tests show Qwen3.5-9B and Gemma4-26B-A4B fail end-to-end IoT replication at 0.39 tokens per second

Anthropic evaluated agents built on Qwen3.5-9B, Gemma4-26B-A4B, and GLM-5.2 running reverse-shell tooling against Android TV targets. The four-stage chain required sequential hardware-specific decisions for binary selection, dependency resolution, and platform debugging. Failures at any serial step halted propagation. Performance data recorded 0.39 tokens per second on distributed Android TV boxes, producing 29-51 hour setup windows for Qwen3.5-9B and 71-150 hours for Gemma4-26B-A4B.

Mirai demonstrated static payload delivery at scale in 2016; these agents instead maintain persistent notes across instances to avoid repeated failed exploit sequences. The approach mirrors documented agent scaffolding in papers on tool-use loops but adds weight transport as the payload itself. Dynamic adaptation remains the unclosed gap: inference server configuration on heterogeneous firmware has no reliable recovery path once a step fails.

Current results indicate small-model agents remain below operational thresholds for worm deployment. Larger GLM-5.2 showed partial reliability, yet the tested cohort did not cross the autonomy line. Operational implication is continued reliance on firmware signing, network segmentation, and inference-server isolation rather than model-size restrictions.

Next measurement point is whether quantized 7-13B agents reach sub-10-hour inference bootstrap on ARMv8 IoT fleets within 18 months under identical tooling.