DeepSeek unveils DSpark, an AI breakthrough that delivers responses up to 85% faster, challenging OpenAI and Google on cost

The AI race is no longer just about building smarter models. It is becoming a race to make those models cheaper and faster to run.

Chinese AI startup DeepSeek has introduced DSpark, a new inference framework that it says can deliver responses up to 85% faster. If those gains hold in production, the technology could let AI providers serve far more users with the same hardware, lowering one of the industry’s highest operating costs at a time when demand for GPUs continues to outpace supply.

The release marks DeepSeek’s latest push to improve AI efficiency rather than model size. It comes as Chinese AI companies shift their attention from training larger foundation models to reducing inference costs, an area that is becoming increasingly important for enterprise deployments and consumer AI services.

DeepSeek’s DSpark

DeepSeek detailed the research behind DSpark in a paper published on GitHub on Saturday. The company said the framework addresses one of the biggest bottlenecks in AI serving: the slow, token-by-token generation process used by today’s large language models. Long responses often leave GPUs waiting between computations, reducing hardware utilization and increasing response times for users.

“We introduce DSpark, a speculative decoding framework that unifies high-throughput parallel generation with adaptive, load-aware verification,” DeepSeek wrote in the paper.

To solve that problem, DeepSeek built DSpark around a speculative decoding framework. Instead of generating and verifying every token sequentially, the system uses a lightweight draft model to predict candidate responses before passing them to a larger model for batch verification. The approach allows more work to be completed in parallel, reducing latency without requiring larger computing clusters.

DSpark goes a step further with a semi-autoregressive generation method that produces small groups of tokens instead of generating each token individually. The framework includes a confidence-based scheduling system that adjusts verification dynamically based on computing demand. During periods of lighter workloads, the system performs more verification to maximize GPU utilization. During peak demand, it reduces verification frequency to keep responses flowing quickly.

According to DeepSeek, the result is response-generation speeds that improve by up to 85%.

“DeepSeek, by adopting what it called a speculative decoding framework, DSpark, said it increased per-user response speeds by up to 85 percent, an efficiency gain that could reduce AI systems’ reliance on larger, more powerful chip infrastructure,” The China Morning Post reported.

DeepSeek’s latest release targets one of AI’s most expensive bottlenecks

The implications extend beyond speed. Huang Yong, a Beijing-based programmer, said higher inference efficiency could significantly reduce the amount of computing infrastructure needed to serve AI applications. In practical terms, a GPU that previously handled roughly 100 user requests could potentially process around 185 requests after the improvement, assuming similar workloads.

The work arrives as AI companies face rising infrastructure costs worldwide. Training frontier models still attracts headlines, yet serving those models to millions of users has become one of the largest recurring expenses for AI providers. Every gain in inference efficiency lowers hardware requirements, electricity consumption, and operational costs.

The timing is especially significant for DeepSeek. The company continues to operate under U.S. export restrictions that limit China’s access to the most advanced AI chips. Improving software efficiency offers one path to delivering competitive performance without relying solely on the newest hardware.

DeepSeek tested DSpark across several open-source models, including Google DeepMind’s Gemma and Alibaba’s Qwen family. The results suggest the framework could extend beyond DeepSeek’s own models and benefit developers seeking faster inference without major infrastructure investments.

DeepSeek’s DSpark

The company has released both the research and the DSpark implementation as open source through GitHub and Hugging Face. The project was developed in collaboration with Peking University.

The next AI war isn’t about training models. It’s serving them faster and cheaper.

The announcement comes as inference optimization emerges as one of AI’s most closely watched battlegrounds. Chinese technology companies are racing to reduce serving costs as AI adoption accelerates across businesses and consumer applications.

Tencent highlighted the same challenge on Friday, describing inference efficiency as a major obstacle to deploying large AI models on less capable hardware. The company said it has invested heavily in engineering techniques, including attention optimization, asynchronous compute communication, and memory caching, to increase output speed.

Xiaomi reached a similar milestone earlier this month. Its AI team said the company’s MiMo-V2.5-Pro-UltraSpeed model can generate more than 1,000 tokens per second, making it among the fastest publicly announced AI inference systems.

For much of the past two years, the conversation centered on who could build the most capable AI model. DSpark raises a different question that could shape the industry’s next phase: Who can deliver high-performance AI at the lowest cost?