The journey from 2.5× to 49.9× — every major release of the Firepoint compression engine.
v26.0July 2026
Real lossless binary STL / PLY geometry compression
Native Binary 3D Compression (STL / PLY)
up to 66× on binary STLmeshes & point cloudsfully lossless
New binary-geometry engine for binary STL and binary_little_endian PLY (point clouds and meshes) — the formats that dominate real CAD, 3D-printing and 3D-scan pipelines.
The float32 vertex records are de-interleaved into per-component streams and byte-plane split, so the highly repetitive exponent / high-mantissa bytes of nearby vertices cluster together and the noisy low bits are isolated. An optional integer-delta stage collapses spatially-coherent scans to tiny repeated differences.
Indexed-mesh de-duplication for binary STL: because STL stores every triangle's three vertices in full, a welded mesh repeats each shared vertex ~6×. The engine dedups vertices by their exact float bit-pattern into a unique-vertex table plus delta-coded index triples — a pure reindexing that is byte-exact lossless and roughly halves the compressed size again on real meshes.
Every candidate transform is tried and the smallest is kept; the whole container is then entropy-coded with the best of Brotli / Zstd / LZMA (plus context-mixing in Archive tier).
Measured, roundtrip-verified lossless on synthetic reference assets: a 2.88 MB binary STL sphere compresses 66.3× (Max) / 59.8× (Fast) vs Brotli-11 at 10.8× and LZMA-9e at 10.4×; a 3.0 MB binary PLY point cloud compresses 25.9× vs LZMA-9e at 2.9× and Brotli-11 at 2.4×.
That is ~6× better than the best general-purpose compressor on the mesh and ~8.8× better on the point cloud — because those tools cannot exploit vertex sharing or per-vertex float structure.
We also prototyped spatial (Morton-order) reordering for point clouds and measured it honestly — it lost to plain delta on coherent scans once the permutation cost is counted, so it was deliberately left out. Only transforms that genuinely win on measured data ship.
Every reconstruction is byte-for-byte exact (headers, property order, endianness, face data and any trailing bytes preserved) and is roundtrip-validated inline before the result can ever be selected; anything that does not parse as a supported binary layout safely falls back to the general engine.
ASCII 3D formats (OBJ / PLY-ascii / XYZ / PTS) continue to use the existing vertex-columnar predictive transform.
v24.1July 2026
Restored Brotli-11 in Apex Max + verified competitive benchmark
Apex Max Brotli-q11 Restore + Competitive Benchmark
+15% avg on realistic databeats Brotli-11 & Zstd-22fully lossless
Fixed an over-aggressive speed cap: Apex Max mode was limiting whole-file Brotli to quality 6–9 on files larger than 32 KB. It now runs full quality 11 up to 2 MB (q9 to 8 MB, q6 beyond) — measured to cost only tens of milliseconds while recovering +13–28% ratio.
Re-benchmarked the shipped engine head-to-head, apples-to-apples on the same files: config/XML 50→61×, plain text 48→62×, CSV 5.4→6.5×, server logs 7.8→8.6×, JSON 18.3→20.4× — a ~15% average gain across realistic file types, all still 100% lossless.
Competitive analysis vs the top general-purpose compressors, all measured on the identical corpus: Firepoint averages ~20% higher ratio than Google’s Brotli-11 and ~40% higher than Meta’s Zstd-22, and wins 5 of 6 realistic categories.
Plain text now flips from a loss to a win vs Brotli-11 (62.2× vs 61.1×); XML/config extends its lead to 61× vs Brotli’s 49× and 7-Zip/LZMA’s 49×.
Compare page and savings calculators updated to these real measured figures; calculator effective ratios remain conservative (at/below measured) so projected savings are never overstated.
Guard bands (q9 above 2 MB, q6 above 8 MB) keep large-file compression fast — no multi-second stalls introduced.
v21.1June 2026
Type-aware per-column delta encoding
Type-Aware Columnar JSON
up to 26× on real JSONinteger-heavy API datafully lossless
Struct-of-arrays (columnar) reversible transform for array-of-objects JSON — clusters same-field values so the entropy coder compresses them far better
New type-aware per-column encoding: integer / timestamp / sequential ID columns are stored as adaptive delta streams, chosen per-column only when it genuinely shrinks the data
Measured ~19–25% additional reduction on top of plain columnar for integer-heavy data — real API records reach 19.8–25.7× and event streams 14.6–20.6× (apex), all fully lossless
Reconstructs the exact original bytes (per-row key order + indentation preserved) and is roundtrip-validated inline before it can ever be selected — non-reproducible inputs safely fall back
Backward compatible: previously compressed columnar files decode identically; floats, strings and mixed columns are never delta-encoded so output stays byte-exact
Automatically tries brotli / zstd / lzma on the columnar representation and keeps the smallest
v19.0June 2026
Ω³ deep-fold + multi-tier architecture
Ω³ Deep-Fold Engine
up to 148× source code*up to 285× text*~4–12× typical