Raspberry Pi 5 LLM benchmarks
Raspberry Pi 5 · 16 GB RAMAll performance measurements are for this device.
Overview
—Capability & performance
Compare speed, memory and task scoresTask measurements & size
Time, tokens and model parametersBenchmark breakdown
Instruction following, tool calling and long promptsVision overview
Answering questions about imagesQuality & performance
Image performance
Detailed measurements
Run progress
Video understanding
Questions about actions and eventsThe question test uses 32 short clips, shown as eight ordered frames without sound or subtitles. These are small public test selections, not full benchmark scores.
Quality & size
Video performance
Three fixed clips, each repeated three timesDo more frames help?
The same 32 questions, with different visual inputQuestion-only checks reveal what can be guessed without seeing the clip. One frame adds a still image; eight frames add visual coverage and action order. This comparison alone does not prove temporal reasoning.
Scores, failed answers & tested setups
Video model progress
Speech & audio input
Transcription and spoken instructionsWER ↓ fewer word errors · Spoken IFEval ↑ better instruction following · error bars: 95% confidence interval
Speech performance
Accuracy & performance
Compare speech models
— = unavailable or outside the model’s tested tasks. Formats show language-model / audio-encoder precision. Speed and memory refer to the Raspberry Pi 5.
Audio run progress
Results table
Snapshot 2026-10-27T19:32:00.807394+00:00. All tested models; a dash means no matching score is available.
| Model | Capability / 100 | Output tok/s | Prompt tok/s | First token · s | Task time · s | Peak memory · MiB | Parameters · B |
|---|---|---|---|---|---|---|---|
| Qwen3.5-0.8BQ4_K_M · 4,096 context · reasoning off | 67 | 15.1 | 93.5 | 4.6 | 20.9 | 1,267 | 0.8 |
| Qwen3.5-4BQ4_K_M · 4,096 context · reasoning off | 76 | 3.4 | 15.6 | 22.5 | 94.8 | 5,691 | 4 |
| Qwen3.5-2BQ4_K_M · 4,096 context · reasoning off | 53 | 7.7 | 47.5 | 9 | 41.7 | 2,713 | 2 |
| Nemotron 3 Nano 4BQ4_K_M · 4,096 context · reasoning off | 78 | 3.5 | 9.2 | 44 | 167.9 | 4,174 | 4 |
| MiniCPM5 2BQ4_K_M · 4,096 context · reasoning off | 71 | 8.1 | 40.4 | 10.4 | 52.5 | 2,901 | 2.5 |
| MiniCPM5 1BQ4_K_M · 4,096 context · reasoning off | 69 | 20.5 | 114.8 | 3.7 | 22.6 | 1,261 | 1.1 |
| Gemma 4 E4B ITQ4_K_M · 4,096 context · reasoning off | 90 | 3.5 | 14.8 | 30.2 | 120.1 | 7,433 | 8 |
| LFM2.5-1.2B-InstructQ4_K_M · 4,096 context · reasoning off | 65 | 16 | 81.2 | 5.3 | 21.9 | 1,547 | 1.2 |
| Granite 4.1 3BQ4_K_M · 4,096 context · reasoning off | 83 | 4.8 | 25 | 16 | 81.8 | 4,441 | 3.4 |
| Qwen3.5-9BQ4_K_M · 4,096 context · reasoning off | 88 | 1.9 | 9.8 | 42.3 | 173.6 | 10,331 | 9 |
| Gemma 4 E2B ITQ4_K_M · 4,096 context · reasoning off | 75 | 7.4 | 31.9 | 14.6 | 62.1 | 4,535 | 5.1 |
| Granite 4.2 3BQ4_K_M · 4,096 context · reasoning off | 60 | 4.9 | 24.2 | 18.1 | 98.6 | 4,351 | 3.7 |
| SmolLM3 3BQ4_K_M · 4,096 context · reasoning off | 65 | 6.1 | 29.1 | 15.1 | 70 | 4,088 | 3 |
| Qwen3-4B-Instruct-2507Q4_K_M · 4,096 context · reasoning off | 80 | 4.3 | 21 | 18.5 | 98.9 | 5,470 | 4 |
| Ministral 3 3B InstructQ4_K_M · 4,096 context · reasoning off | 51 | 5.5 | 26.9 | 19.2 | — | 4,638 | 3.8 |
| Granite 4.2 8BQ4_K_M · 4,096 context · reasoning off | 53 | 2 | 9.2 | 47 | 257.7 | 10,368 | 8.8 |
| LFM2.5-8B-A1BQ4_K_M · 4,096 context · reasoning on | 35 | 9.4 | 35.4 | 10.8 | 52.2 | 10,003 | 8.3 |
| Olmo 3 7B InstructQ4_K_M · 4,096 context · reasoning off | 73 | 2.5 | 11.8 | 34.2 | 164.1 | 12,720 | 7.3 |
| Granite 4.1 8BQ4_K_M · 4,096 context · reasoning off | 78 | 2.1 | 9.3 | 41.6 | 200.7 | 10,197 | 8.8 |
| G9v3 3BQ4_K_M · 4,096 context · reasoning off | 66 | 6.7 | 31.9 | 13.1 | 68.4 | 3,593 | 3 |
| Ministral 3 8B InstructQ4_K_M · 4,096 context · reasoning off | 39 | 2.1 | 9.8 | 49.4 | — | 9,831 | 8.9 |
| Ling 3.0 TinyQ4_K_M · 4,096 context · reasoning off | 83 | 10.1 | 51.8 | 8.8 | 44.9 | 9,246 | 7.9 |
| Ornith 1.5 9BQ4_K_M · 4,096 context · reasoning off | 58 | 2 | 10 | 43 | 194.1 | 10,795 | 9.7 |
| Qwen3-0.6BQ4_K_M · 4,096 context · reasoning off | 72 | 28.4 | 155 | 2.8 | 21.7 | 1,297 | 0.6 |
| Phi-4 Mini InstructQ4_K_M · 4,096 context · reasoning off | 80 | 4.8 | 23.8 | 16.7 | 86.4 | 5,387 | 3.8 |
| Qwen3-1.7BQ4_K_M · 4,096 context · reasoning off | 69 | 10.4 | 56.9 | 7.2 | 42 | 2,596 | 1.7 |
| Qwen3-8BQ4_K_M · 4,096 context · reasoning off | 77 | 2.4 | 10.7 | 36.8 | 170.8 | 9,454 | 8 |
| Qwen3-4BQ4_K_M · 4,096 context · reasoning off | 87 | 4.7 | 22.2 | 18.6 | 101.5 | 5,471 | 4 |
| SmolLM2 1.7B InstructQ4_K_M · 4,096 context · reasoning off | 47 | 10.8 | 50.6 | 8.7 | 46.8 | 2,891 | 1.7 |
| Hermes 3 Llama 3.1 8BQ4_K_M · 4,096 context · reasoning off | 75 | 2.6 | 11 | 37.4 | 164 | 9,781 | 8 |
| OLMo 2 1B InstructQ4_K_M · 4,096 context · reasoning off | 44 | 14.1 | 77.1 | 5.3 | 27.3 | 2,145 | 1.5 |
| LFM2-700MQ4_K_M · 4,096 context · reasoning off | 67 | 24.7 | 130.7 | 3.3 | 14.8 | 1,039 | 0.7 |
| LFM2-1.2BQ4_K_M · 4,096 context · reasoning off | 60 | 16 | 80.2 | 5.3 | 24 | 1,547 | 1.2 |
| Tulu 3.1 Llama 3.1 8BQ4_K_M · 4,096 context · reasoning off | 79 | 2.6 | 11 | 37.1 | 162.4 | 9,348 | 8 |
| LFM2-8B-A1BQ4_K_M · 4,096 context · reasoning off | 71 | 11.7 | 38.2 | 10.9 | 42 | 9,763 | 8.3 |
| LFM2.5-230MQ4_K_M · 4,096 context · reasoning off | 47 | 72.3 | 451 | 1 | 5 | 408 | 0.23 |
| LFM2.5-350MQ4_K_M · 4,096 context · reasoning off | 47 | 50.9 | 266.8 | 1.6 | 7.1 | 567 | 0.35 |
| SmolLM2 135M InstructQ4_K_M · 4,096 context · reasoning off | 27 | 97.6 | 265.1 | 1.7 | 10.2 | 299 | 0.135 |
| SmolLM2 360M InstructQ4_K_M · 4,096 context · reasoning off | 30 | 42.1 | 98.6 | 4.4 | 23.3 | 593 | 0.36 |
| LFM2-350MQ4_K_M · 4,096 context · reasoning off | 47 | 49.5 | 273.8 | 1.6 | 7.7 | 567 | 0.35 |
| Qwen3-VL 4B InstructQ4_K_M · 4,096 context · reasoning off | 89 | 4.8 | 22.2 | 18.4 | 94.6 | 5,489 | 4.4 |
| Qwen3-VL 8B InstructQ4_K_M · 4,096 context · reasoning off | 90 | 2.4 | 10.9 | 36.1 | 164.5 | 9,494 | 8.8 |
| Spark-X2.5 1.7BQ4_K_M · 4,096 context · reasoning off | 65 | 10.5 | 55.2 | 8.4 | 34.5 | 2,378 | 1.7 |
| Spark-X2.5 4BQ4_K_M · 4,096 context · reasoning off | 62 | 4.4 | 21 | 22.2 | 97 | 5,462 | 4.1 |
| Nanbeige4.2 3BQ4_K_M · 4,096 context · reasoning off | 63 | 2.7 | 12.2 | 39.1 | 218.4 | 5,302 | 4.2 |
| Agents-A1 4BQ4_K_M · 4,096 context · reasoning off | 62 | 3.5 | 19.3 | 19.4 | — | 5,612 | 4.5 |
| Llama 3.2 1B InstructQ4_K_M · 4,096 context · reasoning off | 51 | 14.5 | 85 | 5 | 27.9 | 1,796 | 1.2 |
| Llama 3.2 3B InstructQ4_K_M · 4,096 context · reasoning off | 77 | 5.9 | 28.9 | 14.8 | 70.4 | 4,433 | 3.2 |
| Llama 3.1 8B InstructQ4_K_M · 4,096 context · reasoning off | 75 | 2.6 | 11 | 38.9 | 165.7 | 9,809 | 8 |
| Qwen2.5 0.5B InstructQ4_K_M · 4,096 context · reasoning off | 50 | 29.1 | 91.5 | 4.4 | 20.5 | 745 | 0.494 |
| Qwen2.5 1.5B InstructQ4_K_M · 4,096 context · reasoning off | 72 | 11.9 | 61.4 | 6.6 | 32.7 | 2,070 | 1.5 |
| Qwen2.5 3B InstructQ4_K_M · 4,096 context · reasoning off | 76 | 6 | 29.2 | 13.8 | 66.5 | 3,909 | 3.1 |
| Qwen2.5 7B InstructQ4_K_M · 4,096 context · reasoning off | 86 | 2.7 | 11.6 | 35.2 | 155 | 8,948 | 7.6 |
| Spark-X2.6 4BQ4_K_M · 4,096 context · reasoning off | 83 | 2.5 | 19.2 | 23.8 | 114 | 5,938 | 4.1 |
| LFM3-1.7BQ4_K_M · 4,096 context · reasoning off | 48 | 10.3 | 55.8 | 7.4 | 33.7 | 2,285 | 1.7 |
| LFM3-VL-2BQ4_K_M · 4,096 context · reasoning off | 64 | 10.3 | 56.2 | 7.4 | 30.9 | 2,285 | 1.9 |
| K2 Horizon 7BQ4_K_M · 40,960 context · reasoning on | 60 | 2.2 | 10.4 | 36.1 | 229.6 | 14,498 | 9 |
| K2 Horizon 0.9BQ4_K_M · 40,960 context · reasoning on | 54 | 17.8 | 86.1 | 5.1 | 42.3 | 3,470 | 1.1 |
| K2 Horizon 3.7BQ4_K_M · 40,960 context · reasoning on | 65 | 3.9 | 21.1 | 19.2 | 144.8 | 10,356 | 5.1 |
How to read these results
Capability
Average of instruction following, tool calling and finding information in long prompts. Each uses a fixed public test selection. These are not full official leaderboard scores.
Speed and memory
Output and prompt speed average 10 runs. Task time uses six fixed prompts, repeated three times, with up to 128 output tokens. It is not always the time to a complete answer. Peak memory is the most RAM used during these tests.
Comparing models
Compare the same context and reasoning settings. Small score differences may not matter on your tasks. A dash means no matching measurement is available.