The community leaderboard for local LLM speed
Real tokens-per-second numbers from real hardware, filterable across OS · GPU · hosting tool · engine · backend · model. No sponsors, no cloud providers, no synthetic benchmarks.
Top 3 right now
no runs found for qwen2.5-14b-instruct
What we measure
Gen TPS
Generation throughput — output tokens per second during sustained decoding.
Prompt TPS
Prompt-processing speed — how fast the model ingests the input context.
TTFT
Time to first token — latency from request to the first streamed token.
RAM peak · pressure
Peak memory footprint and whether the run crossed memory-pressure thresholds.
Peak package W
Peak whole-package power draw measured at the wall/SoC during the run.
Sustained thermals
Thermal behaviour over a sustained load — does throughput hold or throttle.