In active development
pixel-press
Local-first, offline video compressor with on-device LLM inference
Open-source, local-first desktop video compressor. Uses on-device LLM inference (llama.cpp + SmolLM2-360M GGUF) to translate plain-language intent into FFmpeg encode plans. Built with Tauri (Rust + React), shipping via GitHub Releases for macOS and Linux.
The journey
Built because no existing compressor respects your files — pixel-press runs 100% on-device: FFmpeg runs as a native subprocess, and a tiny local LLM (SmolLM2-360M, GGUF via llama.cpp) turns plain-language intent like “smallest file” or “keep the subs, strip audio” into real FFmpeg plans.
Technical bet: inference runs fully local (no browser WASM, no in-browser LLM) — a Tauri v2 shell bridges a Rust backend to a React + Tailwind UI. An optional rule engine is the default path, so the app works with zero AI downloaded.
CI release pipeline ships DMG and AppImage/.deb for macOS and Linux on tag push. In active development — not yet presented as a shipped product with users.
- →On-device LLM intent parsing with schema-constrained output + rule-engine fallback
- →Zero-copy output handling — encoded files never enter the webview
- →Huge-file smart probe: native path-based encode for multi-GB sources
- →GitHub Actions release pipeline: 4 shipping targets, tag-push triggered