The 680M Agent Bench
Local models as unattended coding agents on a laptop's integrated Radeon 680M, graded blind.
A laptop iGPU shares one memory bus with everything else, so the first question is which models are fast enough to drive an agent at all, and the second is whether the code they leave behind is worth keeping. Seven mixture-of-experts models ran three coding tasks each in a sandbox, 21 runs in all, and a blind review with probes graded every one of them. A third part puts one of those models on a second inference engine, built for mixture-of-experts models, to see whether it makes the agents faster. A fourth adds image input to local models on the same laptop and checks whether a screenshot actually reaches them through the agent. 4 posts so far. The 3090 Agent Bench ran the same kind of models on a desktop GPU.