Boards / Raspberry Pi 5 + AI HAT+ 2

Pi 5 + AI HAT+ 2 (Hailo-10H): LLM benchmarks

A Hailo-10H NPU with its own 8GB RAM on a HAT, running Hailo-compiled models through hailo-ollama.

At a glance

  • Measured runs18
  • Models tested6
  • Fastest run8.7 tok/s
  • Llama 3.2 1B, Hailo-compiled, hailo-ollama · source ↗
  • Sources5

Where to buy

Every run on this board

ModelQuantRuntimeGen tok/sPrompt tok/sTTFT msSource
Llama 3.2 1BHailo-compiledhailo-ollama8.7––source ↗
published 2026
Qwen2.5-Coder 1.5BHailo-compiledhailo-ollama8.06––source ↗
2026-01-20
Qwen2 1.5BHailo-compiledhailo-ollama8.03––source ↗
tested 2026-01-25
Qwen2.5-Coder 1.5BHailo-compiledhailo-ollama7.94––source ↗
tested 2026-01-25
DeepSeek-R1-Distill-Qwen 1.5BW4A8per M5Stack page7.05–660source ↗
accessed 2026-10-17
DeepSeek-R1-Distill-Qwen 1.5BHailo-compiledhailo-ollama6.83––source ↗
tested 2026-01-25
Qwen2.5 1.5B InstructW4A8per M5Stack page6.82–370source ↗
accessed 2026-10-17
Qwen2.5 1.5B InstructHailo-compiledhailo-ollama6.76––source ↗
tested 2026-01-25
Qwen2.5 1.5B InstructHailo-compiledhailo-ollama6.74––source ↗
2026-01-20
DeepSeek-R1-Distill-Qwen 1.5BHailo-compiledhailo-ollama6.72––source ↗
2026-01-20
Qwen2.5 1.5B Instructnot reportedhailo-ollama6.34––source ↗
2026-04-24
Qwen2 1.5BHailo-compiledhailo-ollama5.89––source ↗
2026-01-20
Qwen2.5 1.5Bnot reportedhailo-ollama4.37––source ↗
2026-04-24
Llama 3.2 3BHailo-compiledhailo-ollama2.65––source ↗
tested 2026-01-25
Llama 3.2 3BHailo-compiledhailo-ollama2.6––source ↗
2026-01-20
Qwen2.5-Coder 1.5Bnot reportedhailo-ollama2.07––source ↗
2026-04-24
DeepSeek-R1-Distill-Qwen 1.5Bnot reportedhailo-ollama1.53––source ↗
2026-04-24
Llama 3.2 3Bnot reportedhailo-ollama1.01––source ↗
2026-04-24

Can the Raspberry Pi 5 + AI HAT+ 2 run…

Can it run DeepSeek-R1-Distill-Qwen 1.5B?Can it run Llama 3.2 1B?Can it run Llama 3.2 3B?Can it run Qwen2 1.5B?Can it run Qwen2.5 1.5B?Can it run Qwen2.5-Coder 1.5B?

Have this board? Submit your own run →

Get the monthly board report

New boards, new runtimes, new numbers. One email a month, no hype.