Chinese AI startup DeepSeek has unveiled a new approach to building larger, more capable AI models without needing the most advanced computer chips that US export controls have restricted.

Engram: A Solution to Optimize Memory Usage

The technique, detailed in a technical paper by DeepSeek founder Liang Wenfeng and researchers from Peking University, tackles a fundamental problem: AI models are getting so large that theyโ€™re bumping up against the memory limits of even the best GPUs. DeepSeekโ€™s solution, called "Engram," creates a more efficient filing system that lets the AI store basic facts separately from complex calculations โ€“ freeing up precious computing power for the harder thinking tasks.

Why Chip Memory Matters

Modern AI models need to access vast amounts of information quickly during training and when responding to queries. That requires high-bandwidth memory (HBM) โ€“ specialized, fast-access memory built into advanced GPUs. China faces a significant disadvantage here, with Chinaโ€™s leading memory chip manufacturer ChangXin Memory Technologies remaining several years behind industry leaders like Samsung, SK Hynix, and Micron.

How the Breakthrough Works

Traditional AI models handle everything through computation โ€“ even retrieving simple information. Engram changes this by letting models "look up" foundational facts more efficiently, similar to how humans might consult a reference book. The technique also helps AI handle longer inputs, which remains a major obstacle for deploying AI chatbots as practical assistants.

Results and Prospects

Testing on a 27 billion parameter model showed performance improvements and more computing capacity for complex tasks. Elie Bakouch, a research engineer at Hugging Face, praised the technique for its practical implementation. DeepSeek is expected to release a V4 model with enhanced coding capabilities, coinciding with the first anniversary of its R1 model. The technical paper is expected to receive scrutiny from AI researchers, as DeepSeek has emerged as a prominent example of Chinese AI innovation operating under US export restrictions on advanced semiconductors.