What is gemma-4-31b-uncensored-nvfp4-dflash?

This repository packages a pre-configured container (a self-contained software environment) for running a large AI language model called Gemma 4 31B on NVIDIA's DGX Spark hardware. The container combines two things: a fine-tuned, uncensored version of the model called Deckard Heretic, and a speed-boosting technique called DFlash (speculative decoding), which uses a smaller draft model to predict multiple tokens at once, then verifies them in parallel, roughly tripling single-request speed and more than doubling throughput at high concurrency compared to the baseline. The container runs on vLLM, an engine designed for efficient AI inference, the process of generating text from a model. It exposes an OpenAI-compatible API endpoint, meaning any tool or code that already works with OpenAI's API can point at this container instead, with no code changes. Supported capabilities include chat completions, tool calling, vision and video input, reasoning traces, and structured JSON output. You would use this if you have access to a DGX Spark machine and want to self-host a high-performance, uncensored large language model for coding, math, reasoning, or text tasks. Setup involves downloading the model weights, pulling the Docker image, and starting a container with a few configuration variables such as memory usage and context length. The README includes detailed benchmark tables comparing performance with and without the DFlash speed-up.