Raspberry Pi · Edge AI · 2026
An NPU with its own 8 GB for the Raspberry Pi 5. It runs only Hailo's compiled models, up to about 2B.
from $200 (typical price)
Researched from manufacturer data and public sources; not tested by us.· Last reviewed Oct 11, 2026
Popular models at Q4_K_M with a 32K-token context. Memory needed from real model files; speed is the theoretical ceiling.
This device runs only models compiled by its vendor (1 of our models).
Model weightsContext cache (KV)RuntimeUsable limit (colour = verdict)
| Model | Verdict | Memory | Speed ceiling |
|---|---|---|---|
| Qwen3 1.7B | Runs | — | |
| Qwen3.5 9B | Not supported | — | — |
| Gemma 4 12B | Not supported | — | — |
| gpt-oss-20b | Not supported | — | — |
| Qwen3.6 27B | Not supported | — | — |
| Qwen3.6 35B A3B | Not supported | — | — |
| Llama 3.3 70B | Not supported | — | — |
| gpt-oss-120b | Not supported | — | — |
| Qwen3.5 122B A10B | Not supported | — | — |
| Qwen3.5 397B A17B | Not supported | — | — |
Theoretical maximum for generating text: memory bandwidth divided by the bytes read per token. Real runs usually reach about half to three quarters of it.
Prices are editorial references (launch price and a typical street price on the date shown), not live prices. In 2026 the memory shortage has pushed many prices well above launch.
Needs a Raspberry Pi 5 (sold separately).
Prices are shown only if updated in the last 24 hours and may change; the retailer’s price at checkout applies. We may earn a commission from these links at no cost to you. Affiliate disclosure
Best for: Pi projects that need a tiny offline language or vision-language model next to camera AI.
For Pi projects only. It is not a general local-LLM accelerator.
Hailo's model zoo lists the supported language models; you cannot load an arbitrary model. 40 TOPS is the INT4 figure (20 TOPS INT8).
Source: Manufacturer specifications
How well the popular local-AI tools run on it today (editorial assessment).
Measured results with their source. We never invent speeds.
No measured results yet. The table above shows the theoretical ceiling instead.
Edge AI
Largest popular model at 4-bit: None of our popular models
A 25 W CUDA board that runs 1–4B models: robotics and smart-home projects, not chat at scale.
Runs any small model with CUDA