DeepSeek unveils architecture to cut AI training costsDeepSeek released a research paper on Thursday introducing Manifold-Constrained Hyper-Connections (mHC), a novel architecture designed to improve training stability and scalability for large AI models while minimizing computational costs, with CEO Liang Wenfeng listed as co-author.The mHC framework builds on ByteDance's 2024 hyper-connection architecture by adding a manifold constraint to restore identity mapping properties and reduce memory overhead, with testing on models up to 27 billion parameters showing stable performance without added computational cost.Industry observers expect DeepSeek to launch a new model before Spring Festival in mid-February 2026, as Liang's personal publication of major technical papers has historically signaled upcoming product releases, including last year's R1 model.Source

DeepSeek unveils architecture to cut AI training costs

DeepSeek released a research paper on Thursday introducing Manifold-Constrained Hyper-Connections (mHC), a novel architecture designed to improve training stability and scalability for large AI models while minimizing computational costs, with CEO Liang Wenfeng listed as co-author.

The mHC framework builds on ByteDance's 2024 hyper-connection architecture by adding a manifold constraint to restore identity mapping properties and reduce memory overhead, with testing on models up to 27 billion parameters showing stable performance without added computational cost.

Industry observers expect DeepSeek to launch a new model before Spring Festival in mid-February 2026, as Liang's personal publication of major technical papers has historically signaled upcoming product releases, including last year's R1 model.

Source