Deep Fusion Memory boosts 13B LLM accuracy by 12% at 0.7 ms per token

Researchers introduced Deep Fusion Memory, an architecture that embeds chunk-level parametric memory directly into a language model's weights. Benchmarks show a 13B model answers open-domain questions 12% more accurately while adding only 0.7 ms per token, versus 30–50 ms for RAG.

• DFM adds 0.7 ms per token versus 30–50 ms for RAG
• Open-domain accuracy improved by 12%
• Memory fine-tuning takes ~200 GPU-hours instead of ~5000
• A 100M-chunk bank adds +2 GB of GPU memory