appConfigEntityId

bodyType

customContents

entityId

entityUuid

isVideo

articlePageQuery

article

articleSchemaContent

content

hooksArticlePageAdvertisementConfigArticle

hooksArticlePageFloatingButtonConfigArticle

appConfig

filter

json

liveArticle

Ben Jiang

Chinese artificial intelligence start-up DeepSeek has rolled out a major upgrade to its flagship V4 model aimed at sharply accelerating AI response generation, as competition among Chinese developers increasingly shifts to reducing serving costs and enhancing user experience.
DeepSeek, by adopting what it called a speculative decoding framework, DSpark, said it increased per-user response speeds by up to 85 per cent, an efficiency gain that could reduce AI systems’ reliance on larger, more powerful chip infrastructure.
AI models’ conventional token-by-token output often slowed when responses were lengthy, leading to low utilisation of graphics processing units (GPU) and high user-perceived waiting time, which was a “primary bottleneck in serving AI”, the company said in research published on Saturday.
DeepSeek said the DSpark module accelerated AI response generation – also known as AI inference, which refers to serving a trained model to respond to user queries – by using a lightweight draft model to propose candidate responses and then verifying them in batches with a larger model, speeding up output.
DSpark further refined the approach with a semi-autoregressive generation method, allowing the model to produce small chunks of tokens rather than strictly one at a time.

It also introduced a confidence-based scheduling system that dynamically adjusted how much verification was applied based on computing demand, helping balance speed and output quality.
More frequent checks were applied when computing demand was low to fully utilise the chips, while fewer checks were used when computing demand was high to ensure faster output.
The new technique could reduce the computing resources needed to serve AI systems, according to Huang Yong, a Beijing-based programmer.
For example, with efficiency gains of up to 85 per cent, a single GPU that previously handled 100 user queries could now process about 185, he said.
While DSpark does not enhance an AI model’s general capabilities, it marks DeepSeek’s latest effort to improve AI system efficiency on less powerful chip infrastructure amid tightening US restrictions on China’s access to advanced semiconductors.
DeepSeek tested the framework on several open-source models, including Google DeepMind’s Gemma and Alibaba Group Holding’s Qwen, suggesting DSpark’s enhancements could be applied broadly, with potential applications for companies seeking improved AI performance without significant computing resource investment. Alibaba owns the South China Morning Post.
The company has open-sourced its research of DSpark, a joint effort with the prestigious Peking University, on source code hosting platform GitHub and HuggingFace, the world’s largest online open-source AI community.
The release comes as Chinese AI developers face mounting pressure to make increasingly powerful models cheaper and faster to serve.
While Chinese AI models have been improving their general capabilities, the next battleground has shifted to AI inference optimisation, where companies seek to lower computing costs while handling surging demand from enterprise and consumer users.
The global AI boom has pushed up demand and prices for hardware infrastructure to serve these systems – from GPUs to memory chips – making efficiency gains critical.
DeepSeek’s introduction of DSpark followed comments by Shenzhen-based technology giant Tencent Holdings on Friday that inference efficiency has become a bottleneck to large-scale deployment of AI systems on inferior hardware.
It added it had made a series of engineering efforts – from attention mechanisms, asynchronous compute-communication to memory caching – to improve output speed.
Earlier this month, the AI team of smartphone-to-vehicle giant Xiaomi said it had improved the output speed for its MiMo-V2.5-Pro-UltraSpeed model to generate more than 1,000 tokens per second – one of the fastest output speeds in the industry.


China overtakes US as top workplace for elite AI researchers, study finds

China and the US are racing ahead on AI. Can they manage risks together?

Xi-Trump dinner puts US tech titans in spotlight – what does it mean for China ties?

From Suzhou to AI: Singapore, China eye new playbook for economic ties

Artificial intelligence (AI)

Faster AI, lower costs: DSpark eases inference bottlenecks and chip strain, says DeepSeek

DeepSeek says DSpark delivers faster results and lower costs, marking a breakthough in inference. Photo: Reuters

The new technique could reduce the computing resources needed to serve AI systems, according to a programmer. Shutterstock

‘Digital ID cards’: China moves to regulate AI agents with unified identity system

The new guideline is China’s first national standard focused on AI agent connectivity. Photo: Shutterstock

As China’s tech firms adapt to AI era, workers worry they’ll be ‘optimised’ out of a job

 Illustration: Brian Wang

Alibaba chip unit T-Head triples capital amid AI hardware bet

Alibaba has become more vocal about its chip unit this year and is reportedly working toward spinning it off with an initial public offering. Photo: Shutterstock

DeepSeek hiring spree: Chinese AI firm seeks newcomers as it pursues AGI

DeepSeek founder Liang Wenfeng. Photo: Shutterstock

Alibaba cuts Qwen AI model costs on Qoder coding platform to capture US workday demand

The Qwen and Alibaba logos. Photo: Reuters

The Alibaba logo on a building in Beijing. Photo: Kyodo

Tech

Big Tech

Tech Trends

Tech War

Business

Companies

Start-up unveils speculative decoding framework that speeds up inference by up to 85 per cent amid China’s push to overcome US AI curbs


Start-up unveils speculative decoding framework that speeds up inference by up to 85 per cent amid China’s push to overcome US AI curbs.


Faster AI, lower costs: DSpark eases inference bottlenecks and chip strain, says DeepSeek

.css-1c6uqr6{color:inherit;font-weight:inherit;font-size:inherit;font-family:inherit;line-height:inherit;overflow-wrap:break-word;}Start-up unveils speculative decoding framework that speeds up inference by up to 85 per cent amid China’s push to overcome US AI curbs

Start-up unveils speculative decoding framework that speeds up inference by up to 85 per cent amid China’s push to overcome US AI curbs