Running LLMs inside PDFs, one 5-second token at a time

What it does

llm.pdf is a stunt build that runs large language model inference entirely inside a PDF viewer. It uses Emscripten to compile llama.cpp into asm.js, then injects that JavaScript into a PDF alongside a base64-encoded GGUF model. The result is a single .pdf file that performs token generation without ever leaving your document reader.

The interesting bit

The project hijacks an old PDF JavaScript injection technique to execute the compiled inference engine, turning what is normally a static document format into a weird, self-contained compute container. It is essentially a technical demonstration that the boundary between “document” and “application” is thinner than your PDF reader would like.

Key highlights

• Compiles llama.cpp to asm.js via Emscripten for execution inside PDF JavaScript
• Embeds the entire quantized model into the PDF using base64 encoding
• Supports custom GGUF models via scripts/generatePDF.py
• Targets tiny models around 135M parameters; the README notes these already take roughly five seconds per token
• Extends the “run unexpected things in PDFs” genre, following prior art like DoomPDF

Caveats

• Performance is glacial: 135M parameter models run at roughly five seconds per token, and anything larger is considered “unreasonably slow”
• Only GGUF quantized models are compatible, and Q8 quantization is recommended for best (relative) speed
• This is explicitly a proof-of-concept, not a practical deployment strategy

Verdict

Grab this if you enjoy security-adjacent stunts, Emscripten oddities, or need a party trick to horrify your DevOps team. Skip it if you are looking for an actual production inference engine.