Deepseek Launches Dspark Framework, Boosting LLM Inference Speed by 85%

WHO:

VentureBeat - DeepSeek, a Chinese artificial intelligence start-up, is the organization behind the DSpark framework.

Techmeme - DeepSeek's flagship V4 model is mentioned in relation to the DSpark framework.

WHAT:

VentureBeat - DeepSeek launched DSpark, a new framework designed to speed up large language model (LLM) inference by up to 85%.

VentureBeat - DSpark improved aggregate throughput by 51% for DeepSeek-V4-Flash and by 52% for DeepSeek-V4-Pro during live production tests.

VentureBeat - Per-user generation speedups of 60% to 85% for V4-Flash and 57% to 78% for V4-Pro were reported compared to the prior MTP-1 production baseline.

VentureBeat - DSpark was tested on models including DeepSeek-V4-Flash (284 billion parameters) and DeepSeek-V4-Pro (1.6 trillion parameters).

Techmeme - DSpark was tested on the Gemma and Qwen models.

HOW:

VentureBeat - DSpark employs a method called semi-autoregressive generation, combining speed with awareness of sequence.

VentureBeat - The framework uses confidence-scheduled verification to adjust how much of each draft should be verified based on model confidence and current serving load.

VentureBeat - DSpark allows for speculative decoding, where a draft model suggests several likely next tokens that the larger model checks in parallel.

WHY:

VentureBeat - The DSpark framework addresses the expensive problem of serving large models quickly and efficiently for real users.

VentureBeat - The motivation behind DSpark is to improve the speed of generating tokens during inference without altering the intended output of the underlying model.

(UPDATED June 29, 2026 23:24 UTC)

DeepSeek, a Chinese artificial intelligence start-up, has launched a new framework called DSpark, designed to accelerate large language model inference by up to 85%. During live production tests, DSpark improved aggregate throughput by 51% for the DeepSeek-V4-Flash model and by 52% for the DeepSeek-V4-Pro model, with per-user generation speedups ranging from 60% to 85% for V4-Flash and 57% to 78% for V4-Pro compared to previous benchmarks. The framework utilizes a method known as semi-autoregressive generation and incorporates confidence-scheduled verification, allowing it to efficiently serve large models while maintaining the quality of the output.