DeepSeek open-sources DeepSpec decoder
DeepSeek released DeepSpec, an MIT-licensed codebase for speculative decoding, plus DSpark, a method it says speeds V4 generation 57-85% over standard decoding.
DeepSeek released DeepSpec on June 27, an open codebase for training and evaluating speculative-decoding draft models. It ships under the MIT license with three methods built in: DSpark, DFlash, and Eagle3, plus data-prep tools, training code, and evaluation scripts.
The headline method is DSpark, a speculative-decoding module for the DeepSeek V4 checkpoints. DeepSeek reports it cuts generation latency 60 to 85 percent on the Flash model and 57 to 78 percent on Pro versus standard single-token decoding, and that it improves token acceptance length by 26.7 to 30.9 percent over Eagle3 in the lab’s tests. Speculative decoding works by drafting several tokens cheaply and verifying them in one pass, so a faster, more accurate drafter means more tokens per second on the same GPU.
What it means for you
Inference cost is the line item that decides whether an AI feature is worth shipping. A drafter that squeezes more throughput from existing hardware lowers that cost without a model swap, and an MIT license means you can actually use it in production.
This keeps widening the case that open models are good enough for a lot of operator work, the same momentum behind GLM 5.2 topping the open-weights leaderboard and US developers reaching for cheaper Chinese models.