Demo 3
Level 1, on your device
The first level of a European brain is one that never leaves the house: Ministral 3 running on the box that serves the site. This page says what we expect on a Raspberry Pi 5 with 16GB, how to try it, and lets you compare your own numbers. Nothing on this page is sent anywhere.
Not measured yet. Everything about speed below is an estimate. We have not run Ministral 3 on a Raspberry Pi. When someone does, the numbers go here, with the date and the exact setup.
The estimate
The article's estimate: "the 3B and 8B, compressed to about 2 to 5GB, should run on a 16GB Pi 5, the 8B slowly; on a tablet, the 3B is the realistic one." (A European brain, as read on 20 November 2026)
The file sizes back the first half of that. Mistral's own GGUF files, at the common Q4_K_M compression:
| Model | Q4_K_M file | Q8_0 file | Vision part, if loaded | Speed on a Pi 5 16GB |
|---|---|---|---|---|
| Ministral 3 3B Instruct | 2.15 GB | 3.65 GB | 0.84 GB | Not measured yet |
| Ministral 3 8B Instruct | 5.20 GB | 9.03 GB | 0.86 GB | Not measured yet |
File sizes from the Hugging Face listings of mistralai/Ministral-3-3B-Instruct-2512-GGUF and mistralai/Ministral-3-8B-Instruct-2512-GGUF, 1 GB = 109 bytes, as read on 20 November 2026. These are download sizes, not memory use; memory use will be higher once the context is allocated.
Ministral 3 is released under Apache 2.0, for edge and local use, in 3B, 8B and 14B sizes. (mistral.ai/news/mistral-3, as read on 20 November 2026)
The exact commands
On a Raspberry Pi 5 running a 64-bit Linux. Pick one of the two.
With Ollama
curl -fsSL https://ollama.com/install.sh | sh
ollama pull ministral-3:3b
ollama run ministral-3:3b --verbose "Write the About paragraph for a Chiswick cheese shop's website."
--verbose prints timings after the answer; the "eval rate" line is the generation speed in tokens per second. Swap 3b for 8b to try the larger one. Install command from ollama.com/download/linux; tags ministral-3:3b, :8b and :14b from ollama.com/library/ministral-3, both as read on 20 November 2026.
To let the side-by-side demo call it from a browser, Ollama must allow that page's address. By default it allows only 127.0.0.1 and 0.0.0.0; others are added with OLLAMA_ORIGINS (Ollama FAQ, as read on 20 November 2026). For this page's address that is:
OLLAMA_ORIGINS=https://example.org ollama serve
If Ollama runs as a system service, set the variable in the service instead, as the FAQ describes.
With llama.cpp and Mistral's official GGUF
curl -LsSf https://llama.app/install.sh | sh
llama serve -hf mistralai/Ministral-3-3B-Instruct-2512-GGUF:Q4_K_M --port 8080
This downloads the Q4_K_M file from Mistral's Hugging Face repository and serves an OpenAI-style API at http://localhost:8080/v1. We give --port 8080 because the server's own default port is 9931. Its CORS setting, --cors-origins, allows any page by default. (llama.cpp README and server README, as read on 20 November 2026) The server prints timings, including tokens per second, after each answer.
Not run by us on a Pi. The commands are from each project's own documentation; tell us if one has changed.
Compare your own measurement
Ran it? Put your numbers here to see what they mean for writing a website. This stays on this page: nothing is sent, nothing is saved, and it is gone when you reload.
| Device | Model | Tokens/s | 150-token paragraph | 600-token page | Entered |
|---|---|---|---|---|---|
| Raspberry Pi 5 16GB | Ministral 3 3B and 8B | Not measured yet | Not measured yet | Not measured yet | The article's estimate |