// Lossless Compression Engine  Â·  1.0 · Final release · AES-256 encryption
Just shipped · V22 Three new fronts, all byte-exact: a context-mixing text engine that beats xz -9e by up to 19.5% on real prose (cmix-class); the first lossless H.264 CAVLC reopen (up to 16% on all-intra — the video gap nobody touches); and estate-level cross-file dedup (1.5–22× vs compressing objects one by one).

M.I.K.E.

Mathematically Improved Kompression Engine

Some data isn't compressed. It's a puzzle waiting to be solved.

MIKE asks a different question. Not "how do I pack these bytes?" but "what rule generated them?" When it finds one, it stores the formula instead of the data. That engine is puzzleCompressor — a genuinely new axis in lossless compression, and the world's best lossless compressor for structured, numeric, scientific and float data: it beats zip, rar, 7z, xz and zstd, often by thousands of times. puzzleCompressor is the engine inside MIKE.

// A real middle-out moment. Named for my son Miguel — I told him computers are just math, and the math is worth loving — with a push from my wife Ana, who took one look and said "you're living the Silicon Valley series." And, of course, an affectionate nod to Mike Judge and his brilliance. Pied Piper was fiction. Middle-out was fiction. puzzleCompressor and MIKE are real.

// 256 KB numeric file · lossless · CRC-verified
262,144 B30 B

MIKE recognised an arithmetic sequence and stored its formula — an 8,738× collapse. Best mainstream tool (rar): 444 bytes. Restores bit-for-bit.

Not a lab trick. MIKE does it on real data too: 357 KB of live Binance market timestamps → 38 B37× smaller than rar, 2,383× smaller than 7z. Same formula, real firehose.

// MIKE Score · Weissman-style*
3,024×

*Homage metric: how many times better MIKE's ratio is than zip's on that file. On random data it's ~1× â€” and honest about it.

// A different kind of compression

For thirty years, compression chased one idea. MIKE adds a second.

Every mainstream lossless compressor — from Deflate to zstd — is a variation on a single paradigm: model the byte stream statistically and encode the residual. It's brilliant, it's mature, and it's asymptotically stuck near the data's statistical entropy. MIKE works on a different axis entirely.

The last ~30 years · one axis

Model the output better

Find repeats, then entropy-code what's left. Real, hard-won progress — but all inside the same question: how predictable is the byte stream? The ceiling is Shannon entropy.

LZ77 → DeflateLZMAzstd / brotli
rANS (~2014) was the last genuinely new primitive.
MIKE · a second axis

Recover the rule that made it

If data came from a rule — a counter, a polynomial, a recurrence, a sampled physical process — the shortest description is the rule itself. MIKE recognises the generator and stores the formula. The ceiling isn't Shannon — it's Kolmogorov.

arithmetic → polynomialfloat columnslinear recurrence
generated data: kilobytes → tens of bytes.
The world's best lossless compressor for structured, numeric, scientific and float data — provably, and by orders of magnitude.

Up to 10,000×+ smaller than xz on generated columns · every byte reversible, CRC-verified against zip / rar / 7z / xz / zstd. Then puzzleCompressor cracked already-compressed data — JPEG (8–27%), Motion-JPEG (14–43%), MP3, and the whole DEFLATE family: PNG, ZIP, PDF, Office 20–44% smaller than xz -9 — now extended to any gzip/zlib (GNU gzip included) Snappy (Spark/pandas/Parquet) and LZ4, all via encoder-agnostic parse-capture, byte-for-byte lossless. And now CAVLC-mode H.264 too — we reopen the CAVLC→CABAC gap losslessly (up to 16% all-intra). Only CABAC H.264 & H.265 remain a tie — they’re already at the arithmetic entropy floor, so there we never expand, and say so.

// Watch

Sixty seconds. The whole story.

What MIKE is, why it works, and the receipts.

// Where MIKE stands — at a glance

Best-in-class where the data has math. Honest everywhere else.

Every result below is CRC-verified against zip, rar, 7z, xz and zstd — all at maximum settings. Full tables further down.

Data typeMIKE vs the whole fieldDetail
Scientific / numeric🏆 BESTcounters, sensors, feature tensors, timestamps — beats everyone, up to thousands×
CSV / TSV tables NEW🏆 BEST45–89% smaller than xz -9 — transpose to columns, turn the digits back into numbers so the generators crush them; where plain LZMA just ties zip/rar
Audio (PCM / WAV)🏆 BESTdethrones rar and 7z — ~25% smaller than rar
Text & code NEW🏆 BESTAt --power, a context-mixing engine (bit-history state maps + direct models + 2-stage SSE + match & word models, context-selected mixer) beats xz -9e by up to 19.5% (measured: 19.5% on a real novel, 14.8% on source) — cmix-class — byte-exact. Balanced/Fast still edge 7z/xz. Kept only when it wins the best-of, so it can never lose.
Executables⇔ EQUAL to bestwithin 0.3–0.7% of 7z (BCJ2), beats zip / rar / xz / zstd
PNG · ZIP · gzip · PDF · Office NEW🏆 BEST20–44% smaller than xz -9 on already-DEFLATE files (37–44% on PNG & zip, ~20% on PDF & Office) — inflate, re-deflate byte-exact, let LZMA finish. And now any gzip/zlib — including GNU gzip — via encoder-agnostic parse-capture, byte-exact: DEFLATE lives inside PNG, every zip, PDF/DOCX/XLSX/PPTX/JAR/APK, and the world’s .gz logs.
Snappy · Parquet NEW🏆 BESTEncoder-agnostic Snappy reopen via parse-capture — we record the exact Snappy parse, so Google-C++/Spark/pandas/pyarrow blobs rebuild byte-exact (not just one encoder). Extends into Parquet: parse the footer, reopen each column-chunk page in place. Verified byte-exact on real pyarrow data, and it wins on the tight pages where general compressors get ~0%..
LZ4 (frame & block) NEW🏆 BESTEncoder-agnostic reopen via parse-capture — we record the exact LZ4 tokens, so any encoder’s .lz4 frame or block rebuilds byte-exact. Verified 80%+ smaller on a real frame..
JPEG images NEW🏆 BEST8–27% lossless recompression — where every other tool gets ~0%
Motion-JPEG video NEW🏆 BEST14–43% lossless — per-frame transcode pooled across frames; raw or AVI
MP3 audio NEW🏆 BESTDeep Huffman reopen — decodes each long-block granule's frequency lines and re-codes them with the context-mixing coder, byte-exact. 3–23% past xz depending on the audio pattern (tested so far) — more on tonal/long-block tracks, less on dense broadband; verify-gated to never lose, falls back on tiny/optimised clips.
H.264 video (CAVLC / baseline) NEW🏆 BESTLossless reopen of the CAVLC→CABAC gap — decode each CAVLC I-slice to its syntax symbols, recompress them with our arithmetic context-mixer (CABAC-class), rebuild the byte-exact original. Up to 16% on all-intra CAVLC, ~2% on typical GOP video, verified byte-for-byte.
H.264 (CABAC) / H.265 video— no gainalready arithmetic-coded — at the entropy floor, nobody wins, and we never expand.
// The idea

Predict the data, then store only the surprise.

Compression is prediction wearing a trench coat. MIKE runs three moves per block and keeps whichever wins — spectacular where structure exists, and it simply never expands where it doesn't.

01 · GENERATOR

Find the formula

If a block is an exact sequence — a counter, a ramp, a generated table — store the rule that makes it. A few bytes for any length.

02 · DECORRELATE

Stride-aware delta

For smooth signals — audio, sensor sweeps, gradients — subtract each value from the one a sample ago, collapsing structure into near-zero residuals.

03 · NEVER EXPAND

Best-of, always safe

A plain strong pass is always a candidate, so a block only ever gets smaller — or stays put on truly random data. No lottery, no bloat.

// Want the whole story, with real numbers and where it loses? → How MIKE gets 357 KB to 38 bytes

// The receipts · full benchmark

Every file, every tool. Size, ratio, and the clock — both ways.

Real files, real tools, real round-trips — all six contenders measured the same way and checked byte-for-byte against the original (CRC). Rows are sorted smallest-first, so the winner is always on top. Competitors run at maximum settings (zip -9, rar -m5, 7z -mx9, xz -9e, zstd -19).

// Don’t trust us — run it yourself. Run ./benchmarks/run.sh from the MIKE repo — it regenerates the whole corpus, runs MIKE and every compressor you have, and byte-exact verifies every MIKE result (mike verify = lossless). Latest engine, every number reproducible on your machine

FileOriginalMIKEBest rivalResult
Fibonacci column617,264 B56 Bxz 597,216~10,600× smaller
Cumulative-sum column1,600,000 B56 Bxz 974,712~17,000× smaller
Counter column262,144 B38 Bxz 9,900~260× smaller
CSV table1,770,492 B169,619 Bxz 341,87250% smaller than xz
Prose text · --power738,046 B174,951 Bxz-9e 216,11219% smaller than xz-9e
Baseline JPEG39,756 B31,533 Bxz 39,00021% smaller · xz gets ~0
PNG image12,182 B1,120 Bxz 5,3084.7× smaller than xz
CAVLC H.264 · all-intra, tightxz +7.9%17% smaller than xz
Sensor column400,000 B66,262 Bxz 74,36811% smaller
Optimized JPEG340,899 B340,534 Bxz 339,988~0% — at the floor (honest)
MP3 audio (tiny)25,121 B24,272 Bgzip 24,167~parity — deep reopen falls back on this tiny sample (honest)
MP3 audio (long-block)321,036 B87,162 Bxz 113,508~23% smaller than xz — deep Huffman reopen (tonal → upper bound)
CABAC H.264 / H.265~0%the one tie — already arithmetic-coded

// lossless verified on every row · MIKE picks the best method per file — linear prediction for audio, BCJ2 for executables, entropy transcode for JPEG — and trades time for ratio.

// New · AI & ML datasets

The dataset compressor. On real ML data, MIKE wins.

The AI boom runs on numeric data — feature tensors, training tables, sensor streams, scientific measurements. That is exactly the shape MIKE was built to model, so we pointed it at real, public ML datasets and ran the same honest gauntlet: full round-trip, CRC-checked, all six tools at maximum settings. On the numeric tables and feature tensors, MIKE produces the smallest file of any tool tested — and buries the everyday tools everywhere. Even on already-dense weight and image tensors — where the free build lands within a fraction of a percent of the best — the current Vault engine now wins outright (bert-tiny weights 7.7% past xz -9e); we show it all, unretouched.

// where it pays off: feature stores · dataset versioning (DVC / git-LFS) · cold-storage of training corpora · telemetry & sensor logs · scientific pipelines

Two tiers, both honest. The MIKE rows below are free MIKE 1.0 — every number reproducible with the downloadable pack. MIKE Vault , our enterprise engine, goes further on tabular data (its column transpose turns the digits back into numbers) — flagged under each table where it wins. Vault figures use the full engine and are not produced by the free binary in the pack.

htru2_features.f32644,328 B · float32 feature tensor · feature stores & cached model inputs
ToolSizeRatioCompressDecompress
MIKE484,058 B75.1%0.805 s0.274 s◄ smallest
7z484,268 B75.2%0.120 s0.051 s
xz484,312 B75.2%0.446 s0.058 s
zstd511,647 B79.4%0.237 s0.008 s
rar526,984 B81.8%0.129 s0.021 s
zip534,508 B83.0%0.136 s0.013 s
california_housing.csv1,423,529 B · numeric tabular dataset · the ML training workhorse
ToolSizeRatioCompressDecompress
MIKE 1.0 (free)328,522 B23.1%2.723 s0.251 s◄ smallest
7z328,869 B23.1%0.717 s0.044 s
xz328,916 B23.1%1.252 s0.044 s
zstd358,721 B25.2%1.145 s0.010 s
rar386,199 B27.1%0.430 s0.029 s
zip396,115 B27.8%0.727 s0.020 s
▸ MIKE Vault : 265,300 B (18.6%) — the column-transpose result, 19% smaller than xz -9e. Enterprise engine; the table above is free MIKE 1.0, which the pack reproduces.
htru2.csv1,730,114 B · scientific measurement table · sensor & telemetry logs
ToolSizeRatioCompressDecompress
MIKE676,154 B39.1%3.487 s0.295 s◄ smallest
xz676,280 B39.1%1.584 s0.077 s
7z678,029 B39.2%0.804 s0.069 s
zstd688,652 B39.8%1.624 s0.012 s
rar742,567 B42.9%0.416 s0.030 s
zip778,518 B45.0%0.224 s0.025 s
winequality-red.csv84,199 B · small numeric table · lab results, IoT batches
ToolSizeRatioCompressDecompress
MIKE 1.0 (free)19,894 B23.6%0.316 s0.194 s◄ smallest
7z19,945 B23.7%0.043 s0.010 s
xz19,952 B23.7%0.059 s0.007 s
zstd21,323 B25.3%0.100 s0.006 s
rar22,600 B26.8%0.035 s0.013 s
zip22,872 B27.2%0.054 s0.007 s
▸ MIKE Vault : 18,275 B (21.7%) — column transpose, 8.4% past xz -9e. Enterprise engine; the free MIKE 1.0 table above is what the pack reproduces.
mnist_t10k_images.uint87,840,016 B · uint8 image tensor · raw pixel datasets
ToolSizeRatioCompressDecompress
xz1,354,388 B17.3%10.269 s0.178 s◄ smallest
MIKE1,364,274 B17.4%12.789 s0.388 sMIKE
7z1,367,369 B17.4%3.258 s0.186 s
zstd1,432,706 B18.3%5.916 s0.045 s
rar1,509,151 B19.2%1.256 s0.072 s
zip1,616,686 B20.6%3.290 s0.074 s
▸ MIKE Vault : 1,354,151 B — the current engine (9|EXTREME + compact container) turns the free-tier tie into the smallest file, edging xz -9e. Enterprise engine; the table above is free MIKE 1.0.
bert-tiny_weights.bin17,756,393 B · trained float32 weights · model checkpoints
ToolSizeRatioCompressDecompress
xz16,308,280 B91.8%11.806 s1.717 s◄ smallest
7z16,311,675 B91.9%3.544 s1.490 s
MIKE16,321,839 B91.9%23.098 s1.947 sMIKE
zstd16,411,434 B92.4%9.443 s0.063 s
zip16,424,563 B92.5%1.294 s0.245 s
rar16,504,980 B93.0%3.087 s0.255 s
▸ MIKE Vault : 15,061,164 B (84.8%) — the float byte-plane transform, 7.7% smaller than xz -9e (1.25 MB off a model checkpoint). Where free MIKE ties, the Vault engine wins on trained weights. Enterprise engine; the pack reproduces the free table above.
// Don’t trust us — reproduce it.

Every file above is real and public. We bundled the exact datasets, a one-command benchmark script, and a README explaining what each file is and where it matters. Run it against your own copy of MIKE and any tools you have installed — the script skips whatever’s missing and CRC-checks every round-trip.

Download benchmark pack · 18.9 MB .tgz sha256 2b652a261e727e4754c0b0ec459b4c5d525bc8c6a1185b9da8448876726653d8

$ tar xzf mike-ai-benchmark-pack.tgz && cd mike-ai-benchmark && python3 bench.py

// 6 real public datasets · lossless verified on every row · MIKE wins outright on all four numeric-dataset files; on trained weights nobody beats ~8% and we don’t pretend otherwise

// New · Real-world firehoses

Where this data actually pours out — many times over.

The win isn’t a one-off dataset. It’s the data that machines emit continuously, forever: counters, waveforms, sensor streams. We grabbed real files straight from those firehoses — live market data, a hospital ECG, a telescope frame, a climate model — and ran the honest gauntlet (full round-trip, CRC-checked, all six tools). These are the doors worth knocking on.

// Binance BTCUSDT · 1-min open times (int64)
357,120 B38 B
37× smaller than rar · 2,383× smaller than 7z
// Binance aggTrade IDs · 3.2 MB sequence (int64)
3,200,000 B38 B
221× smaller than rar · a formula, not a file
// PhysioNet MIT-BIH · real ECG waveform
1,950,000 B669,870 B
beats the whole field · 24% past 7z, 72% past zip
Data classReal-world formWhere it gushes 24/7MIKE on a real file
Counters timestamps, sequence & trade IDs, block / packet / event counters observability & metrics (Prometheus/Datadog), DB write logs, market data, network flow records 357 KB → 38 B · 37–221× past rar
Waveforms vibration, ECG/EEG, AC power, RF/IQ, seismic, sonar, audio predictive-maintenance on rotating machinery, medical monitoring, smart-grid PMUs, telecom ECG: smallest of all 6 · 24% past 7z
Sensor streams temperature / pressure / voltage / flow / GPS / accelerometer industrial historians (billions of points), IoT fleets, smart meters, connected vehicles (CAN bus) sensor ramp: 19.2% vs rar 21.6%
Numeric tables & arrays tick bars, FITS / NetCDF / HDF5, feature matrices quant finance, science archives (astronomy, climate, genomics), ML pipelines ties 7z/xz for the lead · buries zip/rar 1.2–2×

// real public files — Binance market data, PhysioNet ECG, NASA FITS, Unidata NetCDF · every row round-trip & CRC-verified · JPEG, MP3, PNG, ZIP, gzip, PDF and now CAVLC-mode H.264 we cracked — only CABAC H.264 & H.265 stay a tie at the entropy floor (see the scorecard above)

// Home turf

On structured data, MIKE beats the entire field.

Shorter bar = smaller file = better. MIKE versus the best result any of zip / rar / 7z / xz / zstd managed on the same file — every number reproducible with ./benchmarks/run.sh, byte-exact verified.

Fibonacci column · 617,264 B · closed-form u64
MIKE56 B
xz -9 · best rival96.8%
// 617,264 B → 56 B, byte-exact — xz barely dents it. ~10,600× smaller than the best rival.
Counter column · 262,144 B · generated
MIKE38 B
xz -9 · best rival3.8%
// 262,144 B → 38 B — ~260× smaller than xz.
CSV table · 1,770,492 B · id·ts·category·price
MIKE9.6%
xz -9 · best rival19.3%
// 50% smaller than xz — reproduce every number with ./benchmarks/run.sh, byte-exact verified.
// New · Estate-level compression

Stop compressing files. Compress the estate.

Every tool on Earth — MIKE included — shrinks one object at a time. But a real customer estate is the same bytes over and over: nightly backups of the same tree, log rotations, vendored libraries copied into a hundred repos, VM images cut from one golden master. Estate mode does content-defined chunking across the whole corpus, stores each unique chunk once, and hands the deduplicated store to the normal MIKE codec. That’s where per-file 30–80% becomes a multiple.

// vs compressing each object alone (S3 / Azure Blob model)
1.5–22×
Object stores compress each object independently. Estate mode shares across all of them.
// vs a single tar | xz — once the corpus > the compressor’s 64 MiB window
+20.7%
Duplication spaced farther than the LZMA window is invisible to xz. Dedup catches it at any distance.
// vendored / scattered identical objects
2.0×
Even inside the window — xz still pays match overhead per copy; dedup erases it for a 4-byte id.
walk corpus → content-defined chunking → global dedup → [ manifest | unique-chunk store ] → MIKE-encode → reconstruct byte-exact
Corpus (all round-trip, byte-exact)vs tar|xz -9evs per-file xzvs raw
Backup snapshots — 15× daily, fits in windowtie15.5×24.8×
Vendored monorepo — shared libs across projects2.0×22.2×191×
Corpus > 64 MiB window — long-range dups1.26×1.52×9.9×

// honest read: when the whole corpus fits inside xz’s 64 MiB window, tar|xz already finds the cross-file duplication itself — there estate ties it, and the win is entirely vs compressing objects independently (15.5×). Estate beats even a single-stream tar|xz once the data outgrows its window or holds scattered identical objects — which is every real GB–TB estate. Byte-exact, verified, never-expand.  mike c -o estate.mike /path/to/corpus picks it automatically when it wins.

// New · VM disks & snapshot backups

The cloud-provider case: VM images and snapshot chains.

EBS snapshots, AMIs, nightly VM backups — the same disk, over and over. We ran the honest gauntlet on a realistic VM disk (OS binaries, logs, package blobs, zeroed free space) and a chain of daily snapshots. All byte-exact, CRC-verified.

Workload (byte-exact)MIKE resultvs former bestvs raw
Single VM disk (mixed image)6.51 MB−4% vs xz -95.1×
8 daily snapshots (2% churn) — vs compressing each alone11.6 MB4.84×21.8×

// dedup is table-stakes (Borg, Veeam, Proxmox Backup do it) — MIKE’s edge is the codec on the deduplicated bytes: the surviving unique data shrinks further through best-of (reopen the DEFLATE/JPEG living inside disks, generative on structured regions) than the plain zstd those tools run on chunks. And on real GB-scale disks — snapshots spaced far beyond any compressor’s window — it wins outright, not just on the per-object comparison. Native backup drivers for Proxmox · KVM/QEMU · VMware · Hyper-V are on the way.

// For datacenter & storage owners

What does the same data cost you at rest? Do the math.

Enter how many terabytes of each kind of data you hold. The reduction beside it is MIKE’s measured average from our tests (editable). At a flat $30 / TB / year all-in for storage, here’s the annual bill MIKE takes off the table.

Storage price
$/TB/yr
Your data by type
File typeStored (TB)MIKE avg (% smaller)
TB%
TB%
TB%
TB%
TB%
TB%
TB%
TB%
Saved per year
$0
 
Total data entered0 TB
Capacity reclaimed0 TB
Effective estate shrink0%
You’d only need to store0 TB
// yearly saving = Σ(TB per type × % smaller) × $/TB/yr. Reduction defaults are MIKE’s measured test averages — logs/text ~65%, numeric/telemetry ~65% (pure counters & formula columns reach 99.99%), CSV/tabular ~80%, Office/PDF ~20%, PNG/JPEG ~15%, MP3 3–23% depending on the audio pattern (deep Huffman reopen, tested so far; ~8% is a representative broadband average), H.264/H.265 video ~8% (CAVLC up to 16%, CABAC ~0%), VM & snapshot backups ~40% (content-addressed dedup). Every field is editable — put in your own. Lossless, always.
// The part nobody else prints

What MIKE is — and honestly isn't.

No lossless compressor can shrink every file; that's a theorem, not a challenge (pigeonhole). Any tool claiming otherwise is selling snake oil. Here's the truth.

✓ The best tool for structured data

Scientific arrays, sensor logs, counters, generated tables, waveforms — MIKE beats zip, rar, 7z, xz and zstd, sometimes by thousands of times.

✓ Provably lossless

Every block round-trips byte-for-byte, verified against five industry tools across the whole corpus. Zero exceptions.

✓ Now wins on compressed data too

We used to say nobody wins on already-compressed media. puzzleCompressor broke that — JPEG (8–27%), Motion-JPEG (14–43%), MP3, and the whole DEFLATE family (PNG/ZIP/gzip/PDF/Office, 20–44% past xz -9), all lossless. Now even CAVLC-mode H.264 (the CAVLC→CABAC gap, up to 16% all-intra). Only CABAC H.264 & H.265 stay a tie — already at the arithmetic floor, where MIKE never expands.

✗ On executables, a hair behind 7z

v0.3.0’s BCJ2 filter pulls MIKE dead-even with 7z on code (within 0.3–0.7%) and beats zip/rar/xz/zstd — but 7z’s range-coded BCJ2 keeps a razor lead. We don’t claim the win we didn’t get. (Audio, though? Dethroned.)

// Origin story

How MIKE was born

Miguel — the original M.I.K.E.
Miguel — the original M.I.K.E.

I’ve long loved Mike Judge’s genius — especially how perfectly he captures the idiosyncrasy of the IT business. Silicon Valley is brilliant (watch it), right up there with Office Space. If you’ve worked in tech, particularly from the ’80s on, it hits differently — because it’s true.

I relate to it more than I’d like. The hunt for venture capital. The adversity. The lawsuits. The endless complexity that drags a creator away from the one thing that matters — creating — and buries him in time-wasting nonsense everyone somehow expects him to be great at too. I’ve watched useless CEOs and mediocre tech win on marketing instead of merit, funded well enough to bury anyone who’s actually better, all while “making the world a better place.” My colleagues started calling me Gilfoyle. They even got me the “Hail Satan & drink coffee” mug. They’re not wrong about the temperament — just the satanism.

It all came together one afternoon. I was teaching math to my 9-year-old son, Miguel — Mike — and he was not having it. He loves Marvel, Transformers, computers and cars (like his dad); math just wasn’t sexy enough. So I told him: you love computers? It’s all math. Robots? Computers plus mechanical engineering. Cars? Mostly mechanical engineering — which is maths and physics, and most of physics is… math. You gotta love math, kiddo.

“You’re so like Gilfoyle — the same sarcasm, the same dark view of humanity, and just as brilliant at tech. And you’re building Newton AI! Gilfoyle built Son of Anton… and look what came out of that.”— Ana, my wife, looking at Newton

That same day, Ana looked at Newton — the sovereign AI I’m building (yes, it’s on this site) — and made that call. So I went back and rewatched Mike Judge’s masterpiece. And somewhere in the middle of Richard explaining why middle-out beats Shannon and every other compression method, I had an epiphany: brilliant tech-BS, Mike — should’ve gone with math. And then puzzleCompressor clicked in my still-awake brain.

I jumped out of bed. Ran the tests overnight, just like Richard. And… VOILA. It’s real. And, much like middle-out, it shines on some files.

So both Mikes — and Ana — left their fingerprints on this. I named it MIKE. I’m not German, and English spells it compressor with a C, but Kompression with a K made the homage too good to ignore.

MIKE is born.

I dedicate this work to my beautiful family, but also, particularly to:
— my son Miguel (Mike), who needs to love maths a little more, like dad;
— my wife Ana, for being my babysitter (she’s way better than Tara, Gilfoyle);
Mike Judge, for the inspired and inspiring work;
Thomas Middleditch, for his brilliant performance and inspiring relentlessness as Richard Hendricks;
— and the brilliant Martin Starr, whose Gilfoyle gave screen life to about 90% of me — although, Martin, I think Gilfoyle would like the Newton page on this same site better… still.
// Platforms — and why

Where MIKE runs.

One Rust engine, every major platform.

LinuxShippedIt’s THE system. Non-negotiable.
WindowsSupportedA first-class, fully-supported build — GUI, CLI, and the MIKE Vault background service, native x86-64.
AndroidShippedThe same Rust engine, in your pocket — and it’s here.
macOSPlannedOn the roadmap — pending partner negotiations.
iOSPlannedOn the roadmap — pending partner negotiations.
AMIGASomedayThe day Commodore ships the Amiga I actually designed. Not one day sooner.
// and it plugs into your cloud storageCompress once on the client — the provider stores & ships the smaller form. Storage, bandwidth, and CPU all saved.
OneDrive · DropboxBetaThe MIKE Vault mirrors compressed files straight into your sync folder — the provider then uploads the compressed form. Transparent, byte-exact, and working today.
SharePoint · M365PlannedAs a Remote BLOB Storage provider (SharePoint Server) or a Microsoft Graph tiering service (Online) — details in the SharePoint section below.
AWS S3 · Azure BlobPlannedSeekable archives + a range-read primitive are built: compress once, then range-GET only the chunks you need — egress billed on the compressed bytes. The SDK backend is next.
// mike-backup — incremental VM & file backup (dedup + the MIKE codec)
KVM · QEMU · libvirtPreviewNative connector built — reads any qemu-img format (qcow2/vmdk/vdi/vhdx); incremental via QEMU dirty-bitmap changed-block tracking. Full image + per-snapshot block-deltas, byte-exact.
Proxmox VEPreviewThe same QEMU dirty-bitmap path — drops in as a backup target beside Proxmox Backup Server, but the deduplicated data comes out smaller (the MIKE codec reopens what zstd can’t).
VMware vSpherePlannedVADP + Changed Block Tracking via the VDDK SDK — a thin connector over the same backup core.
Hyper-VPlannedResilient Change Tracking (RCT) + WMI/VSS on Windows — same trait, same repo.
Live file treesWorkingRsync-style: copy a live filesystem to the backup repo, deduplicated & MIKE-compressed — unchanged files store nothing, restores byte-exact with permissions & symlinks.
// MIKE Folder for QNAP — a share that compresses itself, transparently
QNAP NASPreviewA one-click .qpkg (Intel x86-64 and ARM aarch64 in one package). A FUSE layer makes a shared folder transparently compressing: files you drop in shrink on disk, and any SMB / File Station / NFS download comes back byte-exact, decoded on the fly — same folder, both ways. See the Toolkit below.
// The toolkit — apps, services & drivers

One engine. Many shapes.

The same lossless core, wherever your files live — a desktop app, a one-click viewer, a background service, a backup driver, and a self-compressing NAS share. All byte-exact, all speaking the same .mike.

The app · GUI + CLI

Shrink anything, get it back perfectly

Drag a file or a whole folder — or one command in the shell. Auto-detects content and picks the best model per block; --safe verifies the round-trip before writing, and it never expands. It even carries a native JXL-class JPEG recompressor — a context-mixing DCT-coefficient coder that takes JPEGs ~15% smaller, byte-exact, with zero external tools.

# compress a folder, restore it byte-exact
$ mike c --power -o project.mike ./project
$ mike d project.mike ./restored
$ mike verify bigfile.bin   # prove lossless
New The viewer · mike-view

Open a .mike without unpacking it

Decodes to a private temp folder, shows the file, and wipes the temp on close. Images & text render natively; PDF, video, Word/ODF open in the app you already use. A --web mode serves it to your browser locally — video even streams — and nothing is uploaded.

$ mike-view report.pdf.mike        # desktop viewer
$ mike-view --web holiday.mp4.mike  # in your browser
   decoding locally… http://127.0.0.1:8722
   Ctrl-C: temp wiped
The driver · mike-backup

Incremental VM & file backup

Content-addressed dedup plus the MIKE codec on the surviving unique data — so deduplicated backups come out smaller than the plain zstd/LZMA that Borg, Veeam and Proxmox Backup run. Full image, then per-snapshot changed-block deltas. Every restore byte-exact.

# full image, then daily deltas (only changed blocks)
$ mike-backup backup  /srv/repo mon disk.qcow2 --qemu
$ mike-backup backup  /srv/repo tue disk.qcow2 --qemu
# or rsync-style live file tree
$ mike-backup backup-tree /srv/repo docs ~/Documents
New · Preview The NAS share · MIKE Folder for QNAP

A folder that compresses itself — transparently

Install the .qpkg, share the folder. Files you drop in shrink on disk; anyone who reads or downloads them gets the original, decoded on the fly. No .mike in sight, no extract step — the same folder, both ways.

what users see (the share)   what’s on disk
MikeFolder/report.pdf  3.1 MB   report.pdf.mike  190 KB
MikeFolder/photo.jpg   2.4 MB   photo.jpg (kept as-is)
— SMB download of report.pdf → original bytes, automatic

// The free app is a public download below. Viewer, backup driver and the QNAP package ship for Linux & Windows (QNAP: Intel + ARM in one .qpkg) — every stored byte verified lossless before an original is ever removed. Enterprise builds & support route through HealthTech Portugal.

// M.I.K.E. Vault · managed service   Enterprise

A folder that compresses itself — and opens like nothing happened.

Beyond the free app, MIKE Vault runs as a background service on Linux and Windows — powered by , the latest generation of MIKE’s engine: a self-shrinking archive tier that keeps cold data tiny and hot access instant. Built for teams sitting on terabytes of logs, telemetry and audio — the exact data puzzleCompressor compresses best.

Drop zone · LongTermBackup/

Files land, MIKE shrinks them, the tree stays

Anything copied into the watched folder — in the root or any subfolder — is SAFE-compressed the moment it stops changing (compress → verify the round-trip → then, and only then, remove the original). Your directory structure is preserved; only the bytes get smaller.

quarterly-telemetry.bin quarterly-telemetry.bin.mike
1.2 GB dropped in → kept at a fraction, verified lossless
Instant access · InstantPeek/

Open a .mike and the real file is just there

Double-click an archive (or a program reads it) and the service decodes a fresh copy into InstantPeek/ and hands it to whatever app you’d normally use — the compressed original never moves. The archive stays small on disk; the peek is instant, disposable, and swept automatically.

click report.pdf.mike decoded copy served opens in your viewer
original archive untouched · encrypted vaults unlock with your key
// Enterprise integration · SharePoint & Microsoft 365

Compress inside SharePoint. Decompress before it’s shown.

A SharePoint library is exactly MIKE’s new wheelhouse — Office files (DOCX/XLSX/PPTX are zip), PDFs (FlateDecode), images (PNG/JPEG) and CSV/Excel exports. As a Remote BLOB Storage provider on SharePoint Server, MIKE compresses every document as it’s stored and restores it byte-for-byte the instant anyone opens it — fully transparent, no plugin, no .mike in sight. (On SharePoint Online, a Microsoft Graph service tiers a library the same way.)

A representative document libraryBytesof originalvs xz -9
As stored in SharePoint466,050100%
xz -9 (strong general)177,28838.0%
MIKE tier98,14821.1%−44.6%

// 78.9% of the content database reclaimed, byte-for-byte lossless — vs 62% for xz -9. Per file: DOCX/XLSX 37–48% past xz, PDF 19–25%, PNG 20–31%, CSV 46–48%. Every file restores identical before it’s served.

The M.I.K.E. Vault service is not a public download. Adamantware OÜ has signed a distribution and exclusivity agreement with HealthTech Portugal — please contact them for any further information.
Contact HealthTech Portugal
// Get MIKE

One engine. Every platform.

One Rust core, every platform. GUI (with the splash) or CLI for Linux and Windows, plus the Android APK — pick your poison.

// New in v0.5.0 — a friendlier app: one smart button that knows whether to shrink or restore, live progress (no more frozen window on big files), and a result card that shows exactly where your file landed. Still archives multiple files & whole folders. (all platforms)

🔒 New in v0.6 — AES-256 encryption. Tick Password-protect, pick a password, and your .mike is locked (AES-256-GCM, PBKDF2-SHA256). Share it safely; only the password opens it. — this is the 1.0 Final release.

// Now distributed through HealthTech Portugal — licensing, enterprise builds and support route through them.

// Verify, your call: add --safe and MIKE compresses then immediately decompresses and verifies the result, so you get full round-trip assurance before you ever trust the archive.

New — three speed/ratio modes, everywhere (app · CLI · service): Fast when you want it quick, Balanced by default, or Power to go the deepest — it also tries both compressed containers per block and keeps the smaller — at the expense of speed (mike c --power, or the Mode selector in the app). The Vault service archives cold backups at Power automatically.

// Distribution

MIKE is now distributed through HealthTech Portugal.

Direct downloads are closed. Please contact HealthTech Portugal for pricing and deals — licensing, enterprise builds (the Vault service & cloud integrations), and support all route through them.

Contact HealthTech Portugal — pricing & deals
LINUX · x86-64via HealthTech PT
GUI one-button UI · drag or browse
Download disabled
CLI terminal & scripts
Download disabled
WINDOWS · x86-64via HealthTech PT
GUI one-button UI · drag or browse
Download disabled
CLI terminal & scripts
Download disabled
ANDROID · arm64via HealthTech PT
APK one-button UI · file browser
Download disabled