You can run a real large language model on a twenty-year-old PowerPC Mac. Elyan Labs maintains a port of llama.cpp for PowerPC Mac hardware — Power Mac G4 and G5 towers, PowerBooks, and iMacs — running under Mac OS X Tiger (10.4) and Leopard (10.5). It is not an emulator and not a remote API: quantized transformer inference executes directly on the machine's own PowerPC CPU. If you have ever wanted to run an LLM on an old Mac, this is the toolchain for it.
llama.cpp PowerPCLLM on PowerPC G4 G5 OS X Tiger Leopardbig-endian port AltiVecvintage Mac AI
Modern llama.cpp assumes a little-endian, C++17-era world. PowerPC Macs are big-endian, ship with ancient system compilers, and stopped receiving OS updates in the late 2000s. This project bridges that gap: a maintained source tree and build recipe that compiles llama.cpp-family inference code on Tiger and Leopard, handles big-endian byte order in GGUF model loading, and produces native binaries for the PowerPC G4 and G5. Load a small quantized model and the Mac generates text locally, on silicon Apple shipped in the early 2000s.
Three problems stack up. First, endianness: GGUF model files
and most quantization formats are written little-endian, so weights and headers
must be byte-swapped correctly on load or the model produces garbage. Second,
the toolchain: stock Tiger and Leopard compilers predate the
C++ standards modern ML code assumes, so the port needs careful source-level
compatibility work rather than a simple make. Third,
memory and SIMD: these machines have limited RAM and no modern
vector ISA — but the G4 and G5 do have AltiVec, PowerPC's 128-bit vector unit,
which is genuinely useful for inference math when the code is written for it.
Partly preservation, partly principle. A Power Mac G5 is a capable computer that the industry declared dead; giving it modern inference workloads proves the hardware was never the limit. It is also the same thesis behind our transformer running on a Nintendo 64 and our large-model inference work on IBM POWER8: language models do not require rented cloud GPUs. Inference can run on hardware you physically own and fully control — what we call proof of physical AI. Elyan Labs also runs G4 and G5 machines as miners on RustChain, where vintage silicon earns a proof-of-antiquity bonus, so an old Mac in this lab does real daily work.
The repository includes the patched source, endian fixes, and build notes for Tiger and Leopard. You will need a PowerPC Mac (G4 or G5), a working developer toolchain as described in the README, and a small quantized model — the smaller the better on 1–8 GB machines. Build, point the binary at the model, and the Mac talks back.
Source code on GitHub Citable release (DOI) Elyan Labs home The N64 LLM
By Elyan Labs. Open source. Every CPU has a voice — including the ones Apple retired.