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

DeepSeek released DSpark, an open-source framework designed to accelerate LLM responses through speculative decoding without changing the model’s intended output. The company also published a technical paper, model checkpoints, and DeepSpec tooling under the MIT license, making the system broadly usable for research and commercial deployment.

DSpark targets one of AI’s biggest deployment bottlenecks: the cost and latency of serving large models at scale. Its release reinforces a broader trend toward open infrastructure that can narrow the performance gap between proprietary and open AI systems, especially for chatbots, coding tools, and agentic workflows.