DeepSeek strikes again.
On January 1, 2026, DeepSeek published a new research paper: mHC (Manifold-Constrained Hyper-Connections). It's an improvement to the Transformer architecture that makes training more stable and cheaper.
A new flagship model is likely coming by mid-February.
Their playbook is always the same: first a technical paper, then the model. Liang Wenfeng, the founder, personally published this paper on arXiv (just as he did before for R1 and V3).
As a reminder: one year ago, DeepSeek released R1 and sent Nvidia's stock plunging 17% in a single day (-$600B in market cap)!
DeepSeek continues to play a key role in the new AI paradigm:
- The gains from simple scaling are running out
- LLMs keep improving through architectural enhancements and Reinforcement Learning
- Costs are shifting toward inference (vs. training) but the price of tokens is collapsing (divided by 1,000 in 3 years), making it possible to increase usage without costs exploding
2026 is shaping up to be just as intense as 2025.