Shows how long local hardware takes to serve a realistic agent workload, with 18.7 minutes for Qwen3.8 27B on the RTX Spark laptop, helping decide which machine can run coding agents locally.
Key takeaways · AI-distilled
AA-AgentPerf-Local replays recorded conversations with a fixed output-tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition → count per turn. Its completion time measures model serving and excludes tool execution, rather than timing an AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → solving fresh tasks.
Artificial Analysis uses system-specific runtimes, weight formats and speculative-decoding settings in AA-AgentPerf-Local. The published comparison therefore measures hardware and serving recipes together.
The reported RTX Spark laptop result used a prototype with 128 GB of memory, plugged in and set to high-performance mode. Its 18.7-minute Qwen3.8 27B result does not establish battery-powered performance.
In the reported local benchmarkA standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.Full definition →, Ling 3.0 Flash's 77 GB of weights fit the 128 GB systems but exceeded the tested MacBook's 64 GB and RTX 5090's 32 GB. Available memory limits model choice independently of serving speed.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
