DeepSeek open sources DSpark, speeding up LLM inference by up to 85%

Hugging Face Models

DSpark is a speculative decoding framework that accelerates DeepSeek-V4 per-user generation 60–85% over MTP-1. The open-source release includes checkpoints, training code, and a draft module, no retraining required.

DeepSeekSpeculative decodingInference

See 8 more sources

deepseek-ai/DeepSeek-V4-Flash-DSpark2 days agodeepseek-ai

DeepSeek just made their AI 5x faster. Without changing the model at all. It is called DSpark. And...1 day agoVaibhav Sisinty

DeepSeek open sources DSpark, a new framework to speed up LLM inference by up to 85%10 hours agoCarl Franzen

What Is DeepSpark? How DeepSeek Made Every LLM 50–400% Faster Without Retraining14 hours ago

DeepSeek Releases DSpark, a Speculative Decoding Framework That Accelerates DeepSeek-V4 Per-User Generation 60–85% Over MTP-12 days agoAsif Razzaq

DeepSpec - a deepseek-ai Collection1 day agopmttyjiDiscuss

DeepSeek open-sources inference optimizations with 60–85% faster generation [pdf]via Hacker News2 days agoaurenvaleDiscuss

deepseek-ai/DeepSeek-V4-Pro-DSpark • Huggingface2 days agoExternal_Mood4719Discuss

Hugging Face Models·deepseek-ai

Jun 27, 3:02 AM