miaai-lab/deepseek-v4-flash-dspark-abliterated-uncensored

This project is a set of scripts for running a large language model, an "abliterated" build of DeepSeek V4 Flash DSpark, across two connected NVIDIA DGX Spark computers so it can be queried like a normal AI API from your own machine. Abliterated is a term used in the AI community for a model that has had its built in refusal behavior removed, so it will attempt to answer prompts a standard, safety tuned version of the model would decline. The setup targets a fairly specialized hardware situation: two DGX Spark machines connected together over a fast InfiniBand network, working as a single cluster to serve one very large model. A single script handles the entire launch process from the main machine. It pulls the container image needed to run the model, copies that image and the multi gigabyte model weights over to the second machine, starts the model on both machines in a coordinated way, and waits until the system reports it is ready. Once running, it exposes a familiar chat style API endpoint on a fixed port that other tools can send requests to, the same shape used by many AI chat applications. Behind the scenes the project uses vLLM, a widely used tool for serving large language models efficiently, along with several specific optimizations: compressed number formats to save memory, speculative decoding to speed up responses, and a very large context window that can hold roughly a million tokens of conversation or documents at once. For advanced users, the repository also documents, and provides the code for, the process originally used to strip the model's refusal behavior in the first place. This involves feeding the model both harmful and harmless example prompts, measuring how its internal activations differ between the two, computing a mathematical "refusal direction" from that difference, and then removing that direction from the model's internal weights. The README presents this as something you could rerun yourself with your own prompt sets if you wanted to adjust which layers or how strongly the refusal behavior is altered. Given the specialized dual GPU hardware requirement and the advanced rebuild pipeline, this project is aimed squarely at people already experienced with self hosting large AI models, not casual or non technical users.