You want to try local AI. Does the first step have to be buying a discrete GPU?
An ai-tier test offers a smaller starting point: Spark-X2.5-1.7B Q4_K_M on a Raspberry Pi 5 with 16GB RAM.
The reported output rate is 10.5 tokens/s, with 2,378 MiB peak memory including the loaded model. The test used a 4,096-token context, reasoning off, and ten repetitions of 256 input / 128 output tokens.
But throughput is only part of the experience. The source also reports 8.4 seconds to first text and 34.5 seconds task time in its response test: six fixed prompts, repeated three times.
These are third-party measurements from October 19, 2026 (UTC), not our rerun. Memory excludes other applications; long conversations may behave differently. The result does not establish compatibility with a 2GB Pi.
Start with your real task, then measure the wait as well as the speed.
