DeepMem: byte-level multiscale memory cuts GPU memory 54% with lossless decoding
Afshin Khadangi · hf · 2026-11-04
Afshin Khadangi (SnT, University of Luxembourg) proposes DeepMem, a causal multiscale byte-level architecture combining invertible subband transforms and fast weight memory for sequence compression, with lossless carriers and attention-based super-resolution decoding guaranteeing exact byte identity. Reported gains vs Transformer: BPE perplexity -7.7% (Thai) and -3.1% (FineWeb-Edu), GPU memory -54%, throughput +37%.
More from Infra
- SurrealDB 4 launches as multi-model ACID database for the agentic era — Scobleizer · 2026-11-04
- Fire at Taiwan Panasonic PCB materials plant may tighten CCL supply further — zephyr_z9 · 2026-11-04
- Same Claude model served by Bedrock one day and Anthropic the next, dev finds — georgos_pappas · 2026-11-04
- Self-improving software is inevitable: closing the agent-observability loop — bibryam · 2026-11-04
- Gigabyte's AI TOP motherboard packs 2TB RAM and 9 GPUs for home AI labs — vincent_koc · 2026-11-04
- Building a 7-GPU 132GB VRAM local inference rig: consolidate on EPYC or split across two systems? — infectiousstupidity · 2026-11-04