Brevis Treats Lossless Tensor Compression as Program Synthesis, Cutting Storage by 33%

SingaporeManagementUniversity · hf · 2026-08-06

Singapore Management University introduced Brevis, a novel approach that formulates lossless tensor compression as a program synthesis problem.

• Core Idea: It designs a typed domain-specific language (DSL) capturing tensor structures like repeated regions via reversible operators. Brevis synthesizes a self-contained DSL program for any given tensor to reconstruct it bit-exactly.
• Performance: Tested on 10 public checkpoints across language, audio, and image generation models, it reduced 2.13 TB of checkpoint data to 1.41 TB, achieving a 33.93% storage reduction.
• Advantages: It produces archives up to 30.87% smaller than general-purpose compressors like zstd and gzip, and also beats tensor-specific tools like ZipNN and DFloat11.
• Throughput: Under practical concurrency, it hits 3.60 GB/s compression and 6.61 GB/s decompression speeds without losing a single source byte.

Original post →

More from Infra

• Sam Altman-Backed Oklo Achieves Reactor Criticality in Under a Year — TinfoilTricorn · 2026-08-06
• Run 193B parameter model on 24GB RAM: 8 specialist models with dynamic routing — Similar_Wealth_1850 · 2026-08-06
• Truespar Launches Paddock: High-Performance Local LLM Engine for NVIDIA GPUs — wandedob · 2026-08-06
• Breaking the AI Memory Wall: CXL Moves Toward Commercial Deployment — BenBajarin · 2026-08-06
• Redefining Productivity: 'Intelligence Per Watt' (IPW) as the New Economic Metric — NinaDSchick · 2026-08-06
• Rack-Scale AI Infrastructure Accounts for Only ~10% of Installed Base — BenBajarin · 2026-08-06

Infra channel →