what is gemma4.pas fr?

Gemma4.pas is a complete AI inference engine for Google's Gemma 4 E4B model, written entirely in Delphi and Pascal. It has no dependency on Python, C, or C++ under the hood, and no third-party libraries at all. You load a single packaged model file, called a VPK archive, and call one function to start generating streamed responses from your GPU. The engine covers the full range of what Gemma 4 E4B can do: generating text with an optional private reasoning step before answering, understanding images at different levels of detail, understanding audio from WAV files, and sampling frames from video clips to describe what is happening in them. All of the heavy math, like matrix operations and attention calculations, runs on the GPU through Vulkan compute shaders rather than the CPU. The project also bundles a smaller model called EmbeddingGemma300m for turning text into embeddings, which is useful for semantic search and retrieval tasks, often called RAG. On top of the core engine sits an optional application layer with higher level building blocks: a tool calling system so the model can trigger functions like checking the weather or searching the web, a persistent memory system backed by SQLite that supports both keyword and vector based recall, and a ready made interactive chat interface with slash commands for saving state, adding facts, and restoring past conversations. This project is aimed at Delphi and Pascal developers who want local AI without leaving their existing toolchain, at application developers who want a single file engine with full multimodal support on consumer hardware, and at privacy conscious users who want an assistant that never sends data over the network. To get started, you download the packaged model from HuggingFace, open the included testbed project in Delphi 12 Athens or later, and build it for Win64.