DeepSeek open sources DSpark, a new framework to speed up LLM inference by up to 85%

Even as the geopolitical conversation around AI continues to grow more fraught following the U.S. government's actions to limit the new models from Anthropic and OpenAI, Chinese open source darling DeepSeek is back with yet another open release that could once again change AI development around the globe. Over the weekend, the firm released DSpark, a new, MIT-Licensed system designed to make large language models answer faster without changing what the underlying model is trying to say. The easiest way to think about it is this: most AI chatbots write like someone crossing a river one stepping stone at a time. They choose one small chunk of text, then the next, then the next. DSpark gives the system a scout that runs a few steps ahead, guesses the likely path, and lets the larger model qui…