llmdev.guide debuts real-world inference rankings

This is a useful antidote to spec-sheet theater in the local inference market, especially for agent workloads where sustained throughput matters more than headline TOPS.

• –The site centers on metrics that actually affect day-to-day use: decode speed, prefill speed, power efficiency, and price-performance.
• –It validates reported throughput against a theoretical bandwidth ceiling, which is a good guardrail against inflated claims.
• –The project is open to community submissions, so it can become more useful than a static lab benchmark if contributors keep evidence quality high.
• –The current site looks early-stage, so the real value will depend on how quickly it accumulates trustworthy device data.
• –By focusing on devices that can run at least a 9B model, it targets the segment most relevant to local agent workflows.