Few AI ‘aha moments’ since ChatGPT have matched the early 2025 introduction of DeepSeek-R1, which achieved state-of-the-art performance at a far lower cost and on an open-source basis. Now, the DeepSeek team is once again at the forefront with mHC, a technique to make LLMs smarter than ever. What does this entail and what does it mean for GenAI in 2026?
The DeepSeek team’s research can be read in full here. In short, mHC, the end result of the research, is a new method for an AI model to store and process information more efficiently. A playful analogy is that mHC builds highways within AI models where previously only backroads existed.
As LLMs cannot grow infinitely large but do improve with size, researchers must find ways to make the technology effective at smaller scales. One well-known method is Mixture-of-Experts, where an LLM activates only a portion of itself to generate a response (text, photo, video) based on a prompt. This makes a larger model effectively smaller and faster during operation. mHC promises to be even more fundamental. It offers the chance to increase model complexity without the pain points of the past.
mHC and building highways
Hyper-Connections (the HC in mHC) showed great promise back in September 2024. They allow information to flow through AI models in a more dynamic way. In other words, AI can form complex connections like never before. The problem was that this deeper knowledge quickly led to confusion: add too many Hyper-Connections and the AI model loses stability and stops learning.
mHC (Manifold-Constrained) finally makes this theory workable. By introducing mathematical structure to these complex routes, the data remains stable. This solves an old problem. The current standard, “Residual Connections,” is safe but essentially functions as a simple 1-to-1 conduit. Until now, the only way to make a model smarter was to make it bigger, which consumes computing power. mHC takes a more fundamental approach: instead of building more roads, it makes the roads themselves intelligent. This allows information to move more richly and efficiently through the network without causing congestion.
The next step in AI?
All this promises a major advance in AI. However, we must wait before this progress is applied in practice. We do not know exactly how GPT-5, the Gemini 3 models, and Claude 4.5 are built. OpenAI, Google, and Anthropic could be far ahead of DeepSeek’s latest publication. Yet, DeepSeek-R2 could suddenly emerge as the frontrunner. That depends on its scale and whether these public breakthroughs were already developed secretly elsewhere.
The point is that we don’t know. What is known about mHC will undoubtedly lead to effective smaller models. The combination of Mixture-of-Experts and mHC may result in high-quality LLMs running on local devices or requiring modest hardware. This will likely lead to market volatility again, as seen with DeepSeek-R1. This stemmed from doubts about the sustainability of hyperscale expansion for AI computing clusters, and these doubts may now resurface.
In any case, the true value of mHC remains to be seen. It sounds promising to make AI smarter without drastically increasing computing power. We rarely see such efficiency gains given the massive budgets poured into this technology. AI remains an industry of breakthroughs and stagnation, with specialists now gaining deep knowledge on how to do more with less hardware. The DeepSeek team excels here and remains unique in its openness. So far, this has positively impacted the affordability and accessibility of AI.