Gemma 4 E4B IT vs Qwen3.5-9B on Pi 5
Both measurements share device, quantization, runtime revision, context capacity, generation mode and workload. Capability also requires matching model versions and settings.
| Model | Capability / 100 | Output tok/s | Prompt tok/s | First token · s | Task time · s | Peak memory · MiB | Parameters · B |
|---|---|---|---|---|---|---|---|
| Gemma 4 E4B ITQ4_K_M · 4,096 context · reasoning off | 90 | 3.5 | 14.8 | 30.2 | 120.1 | 7,433 | 8 |
| Qwen3.5-9BQ4_K_M · 4,096 context · reasoning off | 88 | 1.9 | 9.8 | 42.3 | 173.6 | 10,331 | 9 |
What changes with each choice
Gemma 4 E4B IT: capability 90/100, output 3.5 tok/s, prompt processing 14.8 tok/s, observed peak memory 7,433 MiB. Mean task time: 120.1 seconds on the six-prompt workload.
Qwen3.5-9B: capability 88/100, output 1.9 tok/s, prompt processing 9.8 tok/s, observed peak memory 10,331 MiB. Mean task time: 173.6 seconds on the six-prompt workload.
How to use this comparison
Start with observed memory and response speed, then inspect the benchmark category relevant to your task. The capability mean weights instruction following, offline tool calling and context retrieval equally. Small score differences in these benchmark subsets do not establish a reliable quality advantage. Outputs differ in length, and quality scores do not tell you how quickly a model will finish your own task.
These are text measurements. Published image or audio support does not establish corresponding performance on the Pi. Use the individual result links for dates, hashes, per-benchmark scores and configuration details.