Run Falcon-H1 Tiny on Raspberry Pi 1 B+

Confirm the board and the build-host architecture

Use a real Raspberry Pi 1 Model B+ and a separate Docker host. Dockcross documents linux-armv6 for Raspberry Pi, but its published images require an x86_64 host.

• The Pi board has four USB 2.0 ports, Ethernet, microSD and micro-USB power.
• Plan to use Ethernet unless you already own a compatible USB Wi-Fi adapter.
• On Apple Silicon, Docker may need x86_64 emulation; this guide does not claim that emulation path is reproduced.

Flash Raspberry Pi OS Lite 32-bit

Use Raspberry Pi Imager to install the 32-bit Lite image. Preconfigure a hostname, user and SSH, then boot the Pi with Ethernet connected.

Download Raspberry Pi Imager →

Note: The Model B+ has no onboard Wi-Fi. Raspberry Pi Imager can store wireless settings, but they only help if you add a compatible USB Wi-Fi adapter.

Clone and pin the reviewed llama.cpp source

Pin the llama.cpp revision reviewed against the video publication date so later upstream changes do not silently alter the build.

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
git checkout 1ec7ba0c14f33f17e980daeeda5f35b225d41994

Create the ARMv6 Dockcross wrapper

Dockcross provides the Raspberry Pi ARMv6 + VFP2 toolchain. Generate its local wrapper script from the official container image.

Dockcross linux-armv6 documentation →

docker run --rm dockcross/linux-armv6 > ./dockcross-linux-armv6
chmod +x ./dockcross-linux-armv6

Configure a lean ARMv6 build

Disable host-native tuning, shared libraries and OpenMP, then explicitly target ARMv6. The Dockcross image supplies the ARMv6 hard-float/VFP2 toolchain; no NEON target is enabled.

./dockcross-linux-armv6 cmake -S . -B build-armv6 -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=OFF -DGGML_NATIVE=OFF -DGGML_OPENMP=OFF -DGGML_CPU_ARM_ARCH=armv6
./dockcross-linux-armv6 cmake --build build-armv6 --target llama-completion -j2

Note: LocalClaw source-reviewed these flags and the llama-completion target, but did not execute the cross-build because no Docker daemon or Raspberry Pi 1 hardware was available in the validation environment.

Download the recommended legacy Q4_0 model

Start with TII's official Q4_0 file. It is 57.2 MB and uses the simpler legacy 4-bit format highlighted by the creator for ARMv6.

mkdir -p models
curl -L --fail --output models/Falcon-H1-Tiny-90M-Instruct-Q4_0.gguf https://huggingface.co/tiiuae/Falcon-H1-Tiny-90M-Instruct-GGUF/resolve/main/Falcon-H1-Tiny-90M-Instruct-Q4_0.gguf

Boot the Pi and verify ARMv6

Connect over SSH, verify the architecture and create a clean project directory before copying the binary and model.

ssh pi@raspberrypi.local
uname -m
free -h
mkdir -p ~/falcon-h1/models

Note: Continue only when uname -m reports armv6l. Replace pi and raspberrypi.local with the user and hostname configured in Raspberry Pi Imager.

Copy the binary and GGUF to the Pi

Run these commands from the llama.cpp directory on the build host, replacing the SSH destination if needed.

scp build-armv6/bin/llama-completion pi@raspberrypi.local:~/falcon-h1/
scp models/Falcon-H1-Tiny-90M-Instruct-Q4_0.gguf pi@raspberrypi.local:~/falcon-h1/models/

Run a constrained offline completion

On the Pi, use one thread, a 128-token context, a 32-token output cap and disabled memory mapping to reduce pressure on its 512 MB, 32-bit environment.

cd ~/falcon-h1
chmod +x ./llama-completion
./llama-completion -m models/Falcon-H1-Tiny-90M-Instruct-Q4_0.gguf -p "Hello, how are you?" -n 32 -t 1 -c 128 --no-mmap

Note: A generated answer proves inference, not correctness. The creator showed factual mistakes even with the larger Q8_0 file.