DeepSeek Open-Sources DSpark Framework, Boosting LLM Inference Speed Up to 85%

Executive Briefing

• Releases DSpark under MIT license via GitHub and Hugging Face, enabling broad commercial and research use
• Achieves 60–85% per-user generation speedup on DeepSeek-V4-Flash and 57–78% on V4-Pro over prior baseline
• Employs speculative decoding with a lightweight draft model that proposes tokens for the main model to verify in parallel
• Supports other open-weight model families including Alibaba Qwen and Google Gemma, extending beyond DeepSeek models
• Addresses AI deployment economics by improving hardware efficiency for chatbots, coding assistants, and enterprise systems