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

Chinese AI firm DeepSeek said over the weekend it released DSpark, a new MIT-licensed framework designed to accelerate large language model inference by up to 85 percent without altering model outputs. The system uses a speculative decoding approach in which a lighter draft component proposes likely next tokens for a larger target model to verify in parallel, allowing the model to advance multiple tokens at once when guesses are correct. DeepSeek published a technical paper, model checkpoints and DeepSpec, a codebase for training and evaluating speculative decoding systems, on its public GitHub and Hugging Face pages. The company applied DSpark to its latest open models, DeepSeek-V4-Flash and DeepSeek-V4-Pro, and said the method also works with other open-weight families such as Alibaba's Qwen and Google's Gemma. In production tests, DeepSeek reported per-user generation speedups of 60 percent to 85 percent for V4-Flash and 57 percent to 78 percent for V4-Pro over its prior MTP-1 baseline at matched system capacity. Aggregate throughput improvements reached 51 percent for V4-Flash and 52 percent for V4-Pro at defined service targets. The release aims to address the high cost of serving large models quickly enough for real-time applications.