Under the Hood: Pouli Mou Architecture
Unlike browsers that send page snapshots to external cloud APIs, Mov embeds an IPC-controlled local inference daemon directly within Gecko.
1. Local Process Isolation
Mov's browser chrome manages a companion llama-server process over a secured, ephemeral loopback connection. It automatically spins up when a query is submitted and cleanly evicts from memory after three minutes of inactivity.
2. In-Memory Tab Extraction
When you click Summarize tab or attach + page, Mov reads the parsed document tree of the active tab directly from the rendering process, isolates readable text from ads or tracking scripts, and feeds it into the local model prompt window.
3. Hardware Accelerated Offloading
The engine utilizes Vulkan compute shaders alongside OpenMP multi-threading to offload matrix multiplication directly to your GPU without needing proprietary CUDA runtimes.
4. Offline Document Reading
Attached PDFs are read with pdf.js, the same engine behind the browser's built-in PDF viewer, so nothing extra is installed. Long documents are split into parts that fit the model's context window, summarized part by part, then combined. Questions pull in only the few most relevant passages, so a 300-page PDF answers as fast as a 3-page one. Files are processed in memory and never written to disk.
5. Per-Profile Memory
Memories you save are stored in a small database inside your browser profile folder. Each profile has its own memory, Sync never touches it, and the AI engine itself stays stateless. Only the few memories relevant to your message are added to the prompt, and the reply shows how many were used.
6. Local Image Understanding
Pouli Mou's models can read images as well as text. Pictures are resized to at most 1,280 px and analyzed by a vision module that runs entirely on your machine. On graphics cards with 4 GB of memory or less, that module runs on the processor instead, so normal chat stays at full speed. Image understanding can be switched off on the Memory page to save memory.
7. The Pouli Mou Models
Pouli Mou 4B powers Fast and Deep mode (Deep thinks step by step before it answers), and Pouli Mou 9B powers Extended mode. Both are based on Qwen 3.5 by the Qwen team (Apache 2.0) and are published openly on Hugging Face: pouli-mou-4b and pouli-mou-9b.