Key takeaways Chinese AI startup DeepSeek has published a research paper introducing Manifold-Constrained Hyper-Connections, a new training method aimed at reducing computational and energy costs whil...
The Hangzhou-based company published a technical paper on Wednesday introducing a framework called Manifold-Constrained Hyper-Connections, or mHC.
The method is designed to improve how AI models scale while reducing computational load and energy consumption during training, according to the research paper co-authored by DeepSeek founder Liang Wenfeng.
The publication comes as Chinese AI companies operate under US export controls that restrict access to advanced semiconductors essential for developing cutting-edge AI systems.
These constraints have pushed Chinese researchers toward unconventional architectural and software-level innovations.
Industry analysts have responded enthusiastically to DeepSeek's latest research. Wei Sun, principal analyst for AI at Counterpoint Research, told Business Insider on Friday that the approach is a "striking breakthrough."
Sun explained that DeepSeek combined various techniques to minimize the extra cost of training a model, adding that even with a slight increase in cost, the new training method could yield much higher performance.
The new architecture enables models to share richer internal communication in a constrained manner, preserving training stability and computational efficiency even as models scale larger.
This addresses a persistent challenge in AI development where increased information sharing between model components often leads to instability that can derail expensive training runs.
The research paper, which lists 19 authors with Liang's name appearing last, was published through the open-access repository arXiv and open-source platform Hugging Face.
The team tested mHC on models ranging from 3 billion to 27 billion parameters, building on ByteDance's 2024 research into hyper-connection architectures.
"Empirical results confirm that mHC effectively … [enables] stable large-scale training with superior scalability compared with conventional HC (hyper-connections)," wrote the researchers, led by Zhenda Xie, Yixuan Wei, and Huanqi Cao, according to the South China Morning Post.
The authors noted that the new method incorporates "rigorous infrastructure optimisation to ensure efficiency" and holds promise "for the evolution of foundational models."
DeepSeek's technical papers have historically served as early indicators of upcoming product releases.
The company drew widespread attention across the industry a year ago when it unveiled its R1 reasoning model, which was developed at a fraction of the cost of comparable systems built by Silicon Valley firms.
Since then, DeepSeek has released several smaller platforms, but anticipation is building for its next flagship system, widely referred to as R2 and expected around the Spring Festival in February.
While the new paper does not explicitly mention R2, its timing and the fact that CEO Liang Wenfeng personally uploaded it to arXiv have fueled expectations about its role in future releases.
However, Sun offered a more cautious perspective. "There is most likely no standalone R2 coming," Sun said, suggesting that since DeepSeek has already integrated earlier R1 updates in its V3 model, the technique could form the backbone of a V4 model instead.
The research underscores China's strategic pivot toward efficiency and cost-effectiveness in AI development as US semiconductor restrictions continue to limit access to advanced chips.
Chinese startups are increasingly turning to architectural innovation to narrow the performance gap with global competitors without matching their hardware budgets.
The mHC framework represents an attempt to address training instability and limited scalability while maintaining computational efficiency.
The researchers wrote that "mHC will help address current limitations and potentially illuminate new pathways for the evolution of next-generation foundational architectures."
Sun noted that DeepSeek's approach signals the company's internal capabilities. By redesigning the training infrastructure end-to-end, the company is demonstrating that it can pair "rapid experimentation with highly unconventional research ideas."
She added that DeepSeek can "once again, bypass compute bottlenecks and unlock leaps in intelligence," referencing the company's breakthrough moment in January 2025 with the R1 model.
The publication also reflects the increasingly open, collaborative culture among Chinese AI companies, which have been releasing a growing share of their research publicly.
SoftBank Completes $41 Billion OpenAI Investment, Secures 11% Stake
AI Forecast To Put 200,000 European Banking Jobs At Risk By 2030
Trooper Sues Washington State Patrol Over AI Deepfake Video
Key takeaways OpenAI has signed a multi-year, $10 billion agreement with AI chipmaker Cerebras Systems to secure computing infrastructure. The deal will deliver 750 megawatts of computing power throug...
KEY TAKEAWAYS: xAI implemented restrictions preventing Grok from editing images of real people in revealing clothing after global backlash California Attorney General Rob Bonta launched investigation...
Key takeaways TSMC reported fourth-quarter net income of NT$505.74 billion, beating analyst expectations and setting a new record. Fourth-quarter revenue reached NT$1.046 trillion, up 20.5% year over...
