running a much worse model at a higher latency (since local GPU power is limited) is a worse experience for Zed.