Meta Ships OpenZL 0.3: Much Faster Decompression and Better Compression Than Zstd

# What changed in OpenZL 0.3

Meta released OpenZL 0.3, an update to its format-aware compression framework that focuses on both compression ratio and throughput. This release centers on two technical improvements: a faster native LZ engine and a neural selector that builds numeric compression graphs to pick codec combinations for each input.

# Performance highlights

At equivalent settings, Meta reports OpenZL 0.3 delivers decompression speeds roughly 144% faster than Zstandard (Zstd). The LZ engine itself is about 33% faster than the same engine in OpenZL v0.2. Compression throughput stayed largely the same as v0.2, so the major runtime improvement is in decompression.

# Compression quality

OpenZL 0.3 introduces a Compression Transformer, a neural component that selects which codecs to combine for a given input. Meta reports this selector compresses on average about 35% better than Zstd at its highest compression level (level 19). That is a reported average across the inputs referenced in the announcement.

Short bullets summarizing reported changes:

• LZ engine overhaul: native LZ code received optimizations yielding ~33% higher LZ performance versus v0.2.
• Neural selector / Compression Transformer: chooses numeric compression graphs and codec combinations per input, improving compression ratio vs Zstd -19 on average.
• Decompression speed optimized significantly: overall decompression measured at ~144% faster than Zstd at equivalent settings.

# What didn't change much

Compression speed is reported to be similar to v0.2. The release focuses on compression ratio improvements and major decompression throughput gains rather than faster compression.

# How to get the release

Meta points users to the OpenZL 0.3 release details and downloads on the project's GitHub repository.

# Why this matters for implementations

Faster decompression can reduce CPU load and latency in read-heavy workloads, while better compression ratios reduce storage and bandwidth needs. The neural selector approach also changes how compression choices are made: instead of a single codec and level, OpenZL uses a learned decision process to assemble codec pipelines tailored to the input.

# Practical considerations before adopting

• Benchmarks matter: reported figures are relative to Zstd at comparable settings. Reproduce tests on your datasets and hardware before switching.

# Where to read more

The announcement links to the OpenZL 0.3 release on GitHub for downloads and technical details.