Running TinyLlama 1.1B locally on a PowerBook G4 from 2002. Mac OS 9, no internet, installed from a CD.
A custom C89 platform runs modern AI models on classic Mac OS, achieving 7.3x speedup with AltiVec SIMD.
A developer has successfully created MacinAI Local, a complete local AI inference platform that runs natively on classic Macintosh hardware without an internet connection. Unlike previous retro AI projects that port a single tiny model, this is a ground-up platform written in C89 targeting Mac Toolbox APIs. It is model-agnostic, capable of running models from GPT-2 (124M parameters) to TinyLlama (1.1B parameters) via a Python export script. A key technical achievement is the use of AltiVec SIMD optimization on the PowerPC G4 processor, yielding a 7.3x speedup—reducing inference time from 2.4 seconds per token to 0.33 seconds with Q8 quantization.
The platform was demonstrated on a 2002 PowerBook G4 Titanium with 1GHz CPU and 1GB RAM running Mac OS 9.2.2. It includes a 100M-parameter custom transformer trained on Macintosh-specific text, a disk paging system to handle models larger than RAM, and integration with the Speech Manager for audible responses. The AI can generate AppleScript to automate system tasks like launching apps and managing files, asking for confirmation before execution. The project's next goal is to get the 68040 build running on a 1993 Macintosh LC 575 or Color Classic.
- Custom C89 inference engine achieves 7.3x speedup on PowerPC G4 using AltiVec SIMD, reducing token generation to 0.33 seconds.
- Runs multiple modern models like TinyLlama 1.1B on a 2002 PowerBook G4 with 1GB RAM via disk paging, without internet.
- Features agentic control, generating AppleScript to automate Mac tasks and using PlainTalk for speech synthesis.
Why It Matters
Demonstrates the extreme limits of local AI, proving sophisticated models can run on 22-year-old hardware, reshaping perceptions of computational requirements.