A new demo shows Google's Gemma 4 model running directly in the browser via E2B, converting natural language prompts into Excalidraw diagrams without server processing.
The proof-of-concept loads a 3.1GB quantized version of Gemma 4 into the browser using WebAssembly, enabling real-time diagram generation from text descriptions. Users can prompt the model to create flowcharts, sketches, and visual designs, which render directly in Excalidraw.
The implementation leverages E2B's infrastructure for model optimization and browser compatibility. Running large language models client-side eliminates server latency and keeps data local, addressing privacy concerns with cloud-based AI services.
The demo attracted 114 upvotes and 46 comments on Hacker News, with discussion focusing on practical applications, model performance in browser environments, and memory constraints. While 3.1GB represents a significant download, the approach demonstrates feasibility for running capable models locally on modern hardware.
This development reflects broader momentum toward edge AI, where computational tasks shift from remote servers to user devices. The combination of quantized models and WASM optimization continues reducing barriers to in-browser machine learning.
The yt-dlp project has announced limited and deprecated support for Bun, the JavaScript runtime. The change affects users relying on Bun to run the popular video downloader.
A new perspective on software development emphasizes writing code with future maintainers in mind. The approach prioritizes readability and clarity over clever optimizations.
A Rust implementation of PostgreSQL has reached a major milestone by passing 100% of the database system's regression test suite. The project demonstrates functional parity with the original C-based database.