Skip to content

THE FIELD JOURNAL

Raspberry Pi AI: What a Pi 5 Really Runs, How Fast, and What to Buy

By Local AI Scout · Beginner guide

Hand-drawn illustration of an image recognition window and a chat window both feeding one small AI chip on a desk

What AI can a Raspberry Pi 5 really run?

A Raspberry Pi 5 runs small chat models such as Llama 3.2 3B and Gemma 4 E2B on its own CPU, at about 4 to 9 tokens per second in published tests. For vision, the AI Camera and AI HAT+ handle object detection, and only the AI HAT+ 2 also runs small language models.

This guide follows one reader: a hobbyist with a Raspberry Pi 5 8GB, or about to buy one, who wants a private chat assistant that never touches the cloud, or a camera that spots people and pets at home, and wants to know which add-on is worth buying.

Choose the Pi 5 on its own if:

Choose the AI Camera if:

Choose the AI HAT+ if:

  • You want vision with a camera of your choice, such as Camera Module 3 from $25, or several vision models at once.

Choose the AI HAT+ 2 if:

Which Raspberry Pi AI accessory does what in October 2026?

All four accessories run neural networks so the Pi's CPU does not have to. The boards carry a Hailo NPU (neural processing unit, a chip built for AI math), and the camera has an accelerator in its sensor. Ratings are in TOPS, trillions of operations per second, a peak figure that, as our guide to AI PCs and NPUs explains, does not predict chat speed. INT8 and INT4 (8-bit or 4-bit whole numbers) ratings are not like for like.

AccessoryChip and ratingMemorySpeeds upChat models?Official US price
AI KitHailo-8L, 13 TOPS (INT8)Uses the Pi'sVisionNo$70 at launch; no longer in production
AI HAT+ (13 TOPS)Hailo-8L, 13 TOPS (INT8)Uses the Pi'sVisionNoFrom $70
AI HAT+ (26 TOPS)Hailo-8, 26 TOPS (INT8)Uses the Pi'sVision, parallel modelsNo$110 at launch
AI HAT+ 2Hailo-10H, 40 TOPS (INT4); 20 TOPS (INT8) per Hailo8GB on boardVision, LLMs and VLMs up to about 6B parameters by Raspberry Pi's figureYes, small ones$200
AI CameraSony IMX500 sensor with accelerator16MB flash caches modelsVisionNo$70
Raspberry Pi AI diagram: the Pi 5 CPU runs chat models, Whisper and Piper, the AI HAT+ runs vision only, the AI HAT+ 2 runs vision and small LLMs, and the AI Camera runs vision on its sensor
Chat models run on the Pi 5's own CPU; the HATs and the AI Camera are mainly for vision, and only the AI HAT+ 2 adds small language models.

Is the Raspberry Pi AI Kit still worth buying?

Only if it is cheap secondhand. The AI Kit, a Hailo-8L module on an M.2 HAT+ board, launched at $70 on June 4, 2024 and is now out of production. The 13 TOPS AI HAT+ has the same specs and replaces it.

What does the AI HAT+ do, and does it help chat?

The AI HAT+ comes in 13 TOPS (Hailo-8L) and 26 TOPS (Hailo-8) versions, launched at $70 and $110 in October 2024. It plugs into the Pi 5's PCIe connector, a fast data link, and runs object detection, image segmentation and pose estimation through the camera software. It borrows the Pi's memory, and Raspberry Pi marks LLMs and VLMs as not supported, so it does nothing for Ollama or any chat model. Our AI HAT+ buying guide compares the 13 and 26 TOPS boards.

What does the AI HAT+ 2 add, and is it faster for chat?

The AI HAT+ 2, launched on January 15, 2026, pairs a Hailo-10H with 8GB of dedicated on-board RAM. Raspberry Pi says that lets it run LLMs and VLMs up to about 6 billion parameters; a VLM, or vision-language model, describes images in words. Raspberry Pi calls its vision performance broadly equivalent to the 26 TOPS AI HAT+.

Hailo's ready-made models are far smaller than 6 billion parameters. Raspberry Pi's launch post listed five LLMs of 1 to 1.5 billion parameters, and the 5.1.1 release Raspberry Pi installs tops out at Llama 3.2 3B. Hailo's current Hailo-10H model list adds Qwen3-1.7B, Qwen3-VL-2B and three Whisper speech models, but nothing above 2 billion parameters, and the two Qwen3 models need a newer Hailo runtime than Raspberry Pi ships.

Chat speed is disputed. Jeff Geerling, testing on an 8GB Pi 5, wrote that "the Pi's built-in CPU trounces the Hailo 10H": the CPU was faster on all five models he ran, such as 12.47 against 8.25 tokens per second on Qwen2 1.5B, and the Hailo chip is limited to 3W against the Pi chip's 10W. Tom's Hardware timed two qwen2:1.5b answers and found the HAT quicker both times, 13.58 seconds against 22.93 seconds on the CPU for a general-knowledge question, though every answer was wrong. Both tests date from the January launch. The HAT does free the CPU, and its models run through hailo-ollama, Hailo's Ollama-compatible server on port 8000, not Ollama itself. Our AI HAT+ 2 deep dive covers its models, setup and limits.

How fast does a Raspberry Pi 5 run a chat model?

Speed is measured in tokens per second; a token is a word or part of one. The work falls on the Pi 5's 2.4GHz quad-core Arm Cortex-A76 CPU, and published figures are few. In Jeff Geerling's open benchmark table, whose Pi 5 CPU runs were added between November 2024 and January 2025, Llama 3.2 3B generated 4.61 tokens per second on an 8GB Pi 5 and 4.88 on a 16GB one, and DeepSeek R1 14B managed 1.20 on the 16GB board.

In August 2026, Raspberry Pi and Google measured Gemma 4 E2B on an 8GB Pi 5 at 9 tokens per second with Google's LiteRT-LM runtime, against 4 with llama.cpp, with peak memory use of 1,432MB against 4,406MB. Treat that as the makers' claim: the same post's appendix lists llama.cpp at 7 tokens per second.

Bar chart of published Raspberry Pi AI chat speeds: Gemma 4 E2B 9 tokens per second, Llama 3.2 3B 4.88 and 4.61, DeepSeek R1 14B 1.20, and an N150 mini PC 9.06
A Pi 5 CPU manages roughly 4 to 9 tokens per second on small models, about half of what an Intel N150 mini PC reached on the same 3B model.

At 4.61 tokens per second a 100-token reply takes about 22 seconds. The newest CPU figure I found is from Geerling's January 2026 AI HAT+ 2 review, which ran Llama 3.2 3B at 5.55 tokens per second on an 8GB Pi 5 with Ollama. I found no Pi 5 Ollama figures from Raspberry Pi itself, and none on the October 2026 OS release.

Third-party boards are arriving too. A Raspberry Pi post written with Sixfab and DEEPX reports 4.64 tokens per second on Qwen3-1.7B for Sixfab's 25 TOPS board, built on DEEPX's DX-M1M chip, and 11.60 for DEEPX's DX-M1 chip, which the post does not tie to a Pi board. Both are the makers' own figures.

Which models fit on an 8GB or 16GB Pi 5?

Downloads are quantized: weights are stored as smaller numbers, often 4-bit, to save memory. Ollama's default llama3.2:3b is its 4-bit q4_K_M build, 2.0GB, with the 1B version at 1.3GB. The library also lists gemma3:1b at 806MB and gemma3:4b at 3.4GB, and gemma4:e2b at 4.6GB in the llama.cpp build Ollama prefers on Linux.

On the 8GB board I would stay with 1 to 4 billion parameter models. The 16GB board can load a 14B model, but at 1.20 tokens per second. Our Raspberry Pi LLM guide covers which models fit each board, their speeds and tuning.

How do you run a local chat model on a Raspberry Pi 5?

The quickest route to the private assistant is Ollama, a free, open-source model runner, on the 64-bit Raspberry Pi OS released October 6, 2026, based on Debian 13 Trixie.

  1. Flash Raspberry Pi OS (64-bit). Write it to the microSD card with Raspberry Pi Imager. The desktop version needs 6,432MB of storage and Lite needs 2,936MB.
  2. Fit the cooler and the right power supply. Clip on the Active Cooler and use the 27W USB-C supply; a 5V 3A supply limits USB peripherals to 600mA.
  3. Update the system with sudo apt update and sudo apt full-upgrade -y, as Raspberry Pi's AI software guide does.
  4. Install Ollama. Run the one-line Linux install command; the install script picks Ollama's arm64 build for the Pi's 64-bit Arm chip and creates a systemd background service. Our Ollama install guide covers first errors.
  5. Chat with a model. Run ollama run llama3.2:3b and ask a question. In a second terminal, ollama ps should show 100% CPU in the Processor column.
sudo apt update
sudo apt full-upgrade -y
curl -fsSL https://ollama.com/install.sh | sh
ollama run llama3.2:3b
ollama ps
Raspberry Pi AI setup card with the commands to update Raspberry Pi OS, install Ollama, run llama3.2:3b and check it runs 100% on the CPU
Five commands get a private chat model running on a Pi 5, with no accelerator needed.

Ollama listens only on the Pi by default, at 127.0.0.1 port 11434; set OLLAMA_HOST to reach it from a browser chat app such as Open WebUI on another machine. Prefer llama.cpp? Its build guide compiles it with cmake -B build and cmake --build build --config Release.

How do you start a smart camera project with the AI Camera?

For the reader's second goal, the AI Camera is the shortest path. Its 12.3MP Sony IMX500 sensor runs the network and sends the output tensors to the Raspberry Pi. Raspberry Pi's model zoo for it covers classification, object detection, pose estimation and segmentation.

  1. Connect the camera to the Pi 5 with its camera connector cable.
  2. Install the camera's firmware and models with sudo apt install imx500-all, then reboot.
  3. Run the object detection demo. The first run might take several minutes while the model firmware uploads; after that, boxes appear around recognized objects.
sudo apt update && sudo apt full-upgrade
sudo apt install imx500-all
sudo reboot
rpicam-hello -t 0s --post-process-file /usr/share/rpi-camera-assets/imx500_mobilenet_ssd.json --viewfinder-width 1920 --viewfinder-height 1080 --framerate 30

With an AI HAT+ and a normal camera, the AI software guide installs dkms and hailo-all (hailo-h10-all for the AI HAT+ 2), then runs a YOLOv8 demo with rpicam-hello -t 0 --post-process-file /usr/share/rpi-camera-assets/hailo_yolov8_inference.json. Our Raspberry Pi AI Camera guide goes further, with its bundled models, Python examples and project ideas.

Can a Raspberry Pi 5 be an offline voice assistant?

Yes for simple commands, slowly for open conversation. A voice assistant chains speech-to-text, a language model and text-to-speech, and each step is documented on Pi hardware, though not always on a Pi 5.

For listening, Home Assistant's Speech-to-Phrase, which understands only a fixed set of home-control phrases, takes 150 milliseconds per command on a Raspberry Pi 5. Open-ended transcription with Whisper, OpenAI's speech model, takes around 8 seconds on a Raspberry Pi 4; I found no official Pi 5 figure. whisper.cpp lists Raspberry Pi as a supported platform, and our guide to running Whisper locally compares the runners.

For speaking, Home Assistant says Piper is optimized for the Raspberry Pi 4 and generates 1.6 seconds of voice in a second with medium-quality models; our local text-to-speech guide compares it with Kokoro and others. The slow link is the language model: at 4.61 tokens per second, a 50-token answer takes about 11 seconds. Our offline Raspberry Pi voice assistant guide builds one two ways, with Home Assistant or a Python script.

How much does a Raspberry Pi AI build cost in October 2026?

Memory costs have pushed prices up four times since late 2025. The 8GB Pi 5 launched at $80 in 2023, rose to $95 in December 2025, then went up $30 in February 2026 and $50 in April 2026, to $175. The 16GB board is now $305, and the latest rise, on October 1, 2026, only touched 2GB boards. Raspberry Pi sells through resellers rather than direct, so the store prices below come from PiShop.us, one of Raspberry Pi's official Approved Resellers, checked on October 27 and 28, 2026, before tax and shipping. Raspberry Pi says these resellers abide by its published maximum prices.

PartNeeded forPiShop.us (Approved Reseller), October 27 and 28, 2026Official price
Raspberry Pi 5 8GBEvery build$175.00$175 after the April 2026 rise
Raspberry Pi 5 16GBLarger chat models$305.00$305
Raspberry Pi 5 4GBCamera-only builds$110.00$110 after the April 2026 rise
27W USB-C Power SupplyEvery build$12.95$12 at launch
Active CoolerEvery build$10.95$5 at launch
Official 64GB microSD card with Raspberry Pi OSEvery build$20.95Not listed
AI CameraSmart camera$74.95$70
AI HAT+ 13 TOPSVision with any camera$76.95From $70
AI HAT+ 26 TOPSHeavier vision$119.95$110 at launch
AI HAT+ 2Vision plus small LLMs$200.00$200
M.2 HAT+An NVMe drive$12.00From $12
SSD Kit 256GBNVMe storage$99.00 (out of stock)Not listed
Bar chart of Raspberry Pi AI build costs in October 2026: Pi 5 8GB chat build $219.85, with AI Camera $294.80, Pi 5 16GB chat build $349.85, Pi 5 8GB with AI HAT+ 2 $419.85
A complete chat build on the 8GB board costs about $220, and the AI HAT+ 2 nearly doubles that.

Totals at those PiShop.us prices: the 8GB chat build (board, 27W supply, Active Cooler, 64GB card) is $219.85, or $294.80 with the AI Camera. The 16GB chat build is $349.85, and the 8GB board with the AI HAT+ 2 is $419.85 before a camera. The AI HAT+ 2 launched at $130 on January 15, 2026, went up $50 on April 1, 2026, and is $200 on its product page and in its July 2026 product brief; I found no announcement for the last $20.

The 64GB card is enough, since Raspberry Pi recommends at least 32GB for the desktop OS. The Pi 5 has one PCIe 2.0 x1 interface, used by an NVMe drive on the M.2 HAT+ and by every AI HAT, and Raspberry Pi's AI guide does not cover using both at once.

Raspberry Pi 5 vs an N100 mini PC vs a used laptop for local AI

If chat is the main job, compare first. An N100 or N150 mini PC is a small desktop on a low-power Intel chip. I could not verify a current official price for a mini PC or a used laptop, so the table leaves prices out.

October 2026Raspberry Pi 5 8GBIntel N100 or N150 mini PCLaptop, used or new
Processor4 Arm Cortex-A76 cores at 2.4GHz4 cores, up to 3.4GHz (N100) or 3.6GHz (N150), 6WVaries by model
Memory8GB LPDDR4X (16GB version available)Up to 16GB, one memory channelVaries by model
Llama 3.2 3B, published speed4.61 tokens/s9.06 tokens/s (N150, 16GB)13.10-23.88 tokens/s on three laptops; no used-laptop result
Peak power in that test13.9W26.4W14W-51.1W
Raspberry Pi AI accessoriesAI Camera, AI HAT+, AI HAT+ 2Not designed for itNot designed for it

Speed and power figures come from Geerling's benchmark table. The N150 mini PC was nearly twice as fast as the Pi 5 at nearly twice the peak power, and a Dell XPS 13 using its built-in graphics was about five times faster at a similar 14W. The Pi wins on size, camera connectors and AI add-ons; a mini PC or laptop wins on chat speed. Our home AI server guide covers bigger builds.

How much power does a Raspberry Pi 5 use for AI, and what are the limits?

Raspberry Pi specifies 5V/5A DC power via USB-C and recommends its 27W supply, whose 5.1V, 5A output is 25.5W; with it, the board raises the USB current limit to 1.6A. That is a ceiling: in Geerling's table the 8GB Pi 5 peaked at 13.9W and the 16GB board at 11.9W while generating text.

Running costs are small. Drawing its 13.9W peak around the clock, a Pi 5 would use about 122kWh a year, roughly $22 at the August 2026 US residential average of 18.16 cents per kWh; the N150 mini PC's 26.4W peak works out to about $42. The Arm cores are progressively throttled between 80°C and 85°C, and Raspberry Pi recommends the Active Cooler under an AI HAT.

The honest limits for the reader in this guide:

  • Small models know less: Raspberry Pi says they are not designed to match the knowledge set of large cloud models.
  • Ollama keeps a 4,096-token context window by default, the text the model holds in view, so long documents get cut.
  • One PCIe connector means one AI HAT, and at launch Geerling hit errors running two models at once on the AI HAT+ 2.
  • 16GB lets a 14B model load, not run quickly: 1.20 tokens per second.

Common questions about Raspberry Pi AI

Can a Raspberry Pi 5 run an LLM?

Yes, on its CPU. Published tests show Llama 3.2 3B at about 4.6 to 5.6 tokens per second on the 8GB board, and Raspberry Pi and Google measured Gemma 4 E2B at 9 tokens per second with LiteRT-LM. Models of about 1 to 4 billion parameters suit it best; a 14B model managed only 1.20 tokens per second.

Does the Raspberry Pi AI HAT+ speed up chat models?

No. The AI HAT+, in its 13 and 26 TOPS versions, and the older AI Kit accelerate vision models such as object detection and pose estimation. Raspberry Pi's documentation lists large language models and vision-language models as not supported on the AI HAT+, so chat models still run on the Pi 5 CPU.

Is the AI HAT+ 2 worth it for a chat assistant?

Not for most people. It costs $200, and Raspberry Pi says its own 8GB fits language models of up to about 6 billion parameters, but Jeff Geerling found the Pi 5 CPU faster on every model he tested, while Tom's Hardware timed quicker replies on the HAT with one model. It makes most sense for vision plus a small model.

What is the difference between the Raspberry Pi AI Kit and the AI HAT+?

The AI Kit was a Hailo-8L module on an M.2 HAT+ board, rated at 13 TOPS (INT8) and launched at $70 in June 2024; it is no longer in production. The 13 TOPS AI HAT+ puts the same chip on a single board, and a 26 TOPS version uses the more powerful Hailo-8.

Should I buy the 8GB or 16GB Raspberry Pi 5 for AI?

For most home projects the 8GB board, $175 in October 2026, is enough: it runs 3B chat models and works with every AI accessory. The 16GB board, at $305, can load larger models, but a published test shows a 14B model at 1.20 tokens per second, so the extra $130 mostly buys room, not speed.

Do I need an AI HAT for object detection on a Raspberry Pi?

No. The $70 AI Camera runs detection models on its Sony IMX500 sensor and sends the results to the Pi, and it works with all Raspberry Pi models. An AI HAT+ makes sense on a Pi 5 when you want a different camera, bigger vision models or several networks running at once.

My verdict for a Pi 5 owner: use the CPU for chat, add the AI Camera for vision

For the reader in this guide, the 8GB Pi 5 with the 27W supply, Active Cooler and a 64GB card, $219.85 in October 2026, is the whole private chat assistant: Ollama and a 3B model, no accelerator. For a smart camera, add the AI Camera ($70 list, $74.95 at PiShop.us), the cheapest and simplest vision add-on. I would not buy any accelerator to speed up chat, and would only pick the AI HAT+ 2 if you need vision and a small language model on one board while the CPU does other work. If chat speed matters more than size, an N150 mini PC ran the same 3B model at nearly twice the Pi's speed in Geerling's table.

Your next step: flash Raspberry Pi OS (64-bit), install Ollama, run ollama run llama3.2:3b, and ask it something you can check, so you know how about 5 tokens per second feels before you buy anything else.

Sources

Keep exploring: Apps · Models · All guides

THE SCOUT LETTER

Good finds.
Less noise.

Useful local AI apps, practical guides and the updates worth your time. A thoughtful note in your inbox.

Free to join. Unsubscribe with one click.