← all series

Hutter Prize compression: enwik9

• Domain: algorithms
• Role: discovery series
• Metric: total size in bytes of decompressor plus archive for a fixed 1 GB text corpus, under a CPU-time and memory cap, per awarded record
• Coverage: 2019 baseline to 2026; the prize moved to enwik9 on 2020-02-21; prize site read 2026-10-03, benchmark page's own update dated 2026-09-30; rows through the 2026-09-21 snapshot date
• Data: enwik9-records.csv, holding both the hutter_enwik9 and ltcb_enwik9 series
• Upstream: http://prize.hutter1.net/ and http://mattmahoney.net/dc/text.html
• Verdict: accelerating — 3 awarded records in 2026 against 0 in 2025 and 4 over 2021–2024; the 2026 awards cut the record 9.4% against 5.0% over 2019–2024

Hutter Prize enwik9 record

Definition

The task is to compress the first 10^9 bytes of a fixed XML dump of English Wikipedia, scored on the compressed size including the size of the decompression program. The corpus was frozen in 2006, so the task has no benchmark drift by construction, and counting the decompressor closes the loophole of hiding the model in the program.

The prize caps resources:

"must run in ≲50 hours using a single CPU core and <10GB RAM and <100GB HDD on our test machine" — Hutter Prize rules, prize.hutter1.net, read 2026 [@hutter2026prize]

The cap excludes GPUs at run time, so two series are plotted rather than one: the prize is the constrained series, and Matt Mahoney's Large Text Compression Benchmark is the same corpus with no cap, admitting GPU and TPU compressors [@mahoney2026ltcb].

A "discovery" is an awarded record, dated by the date the prize site's record table gives it, which is the submission date; the award itself can come months later. The prize pays only for improvements of at least 1% over the standing record, so the ledger records steps that cleared a hurdle, not every improvement made.

Facts

• baseline: 116,673,681 bytes at the 2019 phda9v1.8 baseline
• awards to 2024: starlit by Artemiy Margaritov on 2021-05-31 · fast cmix by Saurabh Kumar on 2023-07-16 · fx-cmix by Kaido Orav on 2024-02-02 · fx2-cmix by Kaido Orav & Byron Knoll on 2024-09-03 at 110,793,128 bytes
• awards of 2026: cmix-lex by Ibrahim Marcouch & Kaido Orav on 2026-06-26 · cmix-obias by David Freelan on 2026-07-19 · fx2-cmix-transformer by Vladimir Ivanov on 2026-07-24 at 100,424,672 bytes; all three awarded together on 2026-09-30
• steps: measured against the record each displaced, the steps are 1.13%, 1.04%, 1.38%, 1.59%, 1.01%, 1.05% and 7.46%; the four awards of 2021–2024 take the total down 5.0% from the 2019 baseline, the three awards of 2026 take it down a further 9.4%, and the standing record is 13.9% below the 2019 baseline
• pending: zmix 1.0 by James Byrne, listed on the benchmark page on 2026-09-14 at 99,312,424 bytes of archive plus compressor, a further 1.11% and inside the 99,420,425 needed to clear the 1% hurdle; undergoing prize testing as of the benchmark page's 2026-09-30 update
• uncapped: nncp v3.2 reached 107,261,318 bytes on 2023-10-23 and stood until 2026-07-24; zmix 1.0 reached 96,096,261 bytes on 2026-09-14, the lowest uncapped entry dated on or before the 2026-09-21 snapshot

The 2026-07-24 step replaced a hand-built model with a small pretrained one:

"It is derived from fx2-cmix by replacing the LSTM model with a transformer with 6M parameters pre-trained on enwik9 on 8 RTX 5090 GPUs for 26 hours, adding 2.9 MB to the compressor and decompressor. It does not need a GPU to run." — Large Text Compression Benchmark, fx2-cmix-transformer entry, read 2026-10-03 [@mahoney2026ltcb]

The award announcement:

"Ibrahim Marcouch & Kaido Orav are the ninth Winners! David Freelan is the tenth Winner! Vladimir Ivanov is the eleventh Winner!" — Hutter Prize, prize.hutter1.net, news banner, read 2026-10-03 [@hutter2026prize]

The retired 100 MB enwik8 prize is a different corpus and is not joined to the curve. Its complete five-row chronology:

Alexander Rhatushnyak set all four enwik8 awards; the gap before the last runs eight and a half years. The prize moved to enwik9 in February 2020.

The collection-wide cumulative index redraws this series as the standing record's value over time:

Method

The rows are transcribed by hand from the two upstream pages, so there is no fetcher that rebuilds them. fetch.py is a staleness probe instead: it reads the prize page's record table and reports if its top awarded row is smaller than the CSV's standing record, leaving the CSV and this document to be updated by hand. check.py recomputes the fact lines above from the CSV.

Prize rows carry the totals the prize site scores, which add the compressor to the self-extracting archive; the pending row carries the same sum as the benchmark page lists it. Benchmark rows carry the benchmark's own total and keep only entries that set a new uncapped low.

figure.py calls the shared compression_chart() in ../../lib/families.py, which reads enwik9-records.csv, keeps the rows whose series column matches, and plots total_bytes divided by 10^6 against the year fraction of date as a step function. Rows whose award column is anything other than pending are drawn filled and joined by the solid line, which therefore includes the unawarded 2019 baseline; the pending row is drawn as an open marker labelled from its program column, "zmix 1.0, pending". When the series is hutter_enwik9, the function additionally selects the ltcb_enwik9 rows and draws them in grey dashes, with a corner note reporting the last uncapped row's size and month. The axis is linear and in megabytes; January 2026 onward is shaded, as in every figure here.

Limitations

• a hurdle censors the small steps. The prize pays only above 1%, so improvements below that threshold are invisible here whether or not they happened, and the 1.0–1.6% steps before 2026-07 are partly an artefact of the rule.
• neural is not AI-authored. fx2-cmix-transformer and the uncapped entries below it carry a transformer trained offline, nncp has used a transformer since 2021, and cmix has carried an LSTM since 2016; a trained model inside a compressor is not a model-written compressor.
• training compute sits outside the cap. The cap binds the run on the test machine; the 26 GPU-hours on 8 RTX 5090s that trained the fx2-cmix-transformer weights are not counted, only the weights' bytes.
• the snapshot date leaves out a later entry. The benchmark page lists cmix-lex-transformer lexth11c, dated 2026-09-27, below zmix 1.0 at 95,836,613 bytes; it postdates the 2026-09-21 snapshot and is not in the CSV.
• the two pages do not agree on every byte count. fx2-cmix is 110,793,128 on the prize site and 110,351,665 on the benchmark, because the prize adds the compressor and the benchmark does not, and the benchmark's nncp v2 row prints a total equal to its archive-only size, apparently omitting a 99,671-byte decompressor.
• the rows were transcribed by hand from prose pages. Neither upstream is a feed; both are HTML tables read and typed into the CSV, and the prize site was read over plain HTTP because its TLS certificate has expired.
• the capped and uncapped series are not comparable in level. They are different rules on the same corpus, which is why they share an axis but not a line style.
• the 2026 awards came out together. The three 2026 records were dated by submission across four weeks but awarded on one day, 2026-09-30, after a 2025 with no award; the dating rule spreads one announcement across three dates.

AI attribution

No record on either the capped or the uncapped series credits a language model or an agent as the author of a compressor, as of the prize-site read of 2026-10-03 and the benchmark page's update of 2026-09-30; entry write-ups and source repositories were not read. The nearby AI-credited results in this domain are not compression records: the 23% GPU kernel speedup reported in the AlphaEvolve paper [@novikov2025alphaevolve], and GPU kernels found by test-time training that beat the best human submissions by 15 to 51% [@yuksekgonul2026learning].

Sources

• [@hutter2026prize] — the prize site: the record table and award dates, the resource cap quoted above, the 1% improvement hurdle, and the award announcement.
• [@mahoney2026ltcb] — the Large Text Compression Benchmark: the uncapped rows, the pending zmix 1.0 entry and its compressor size, the fx2-cmix-transformer description, the lexth11c entry named in Limitations, and the enwik8 chronology.
• [@novikov2025alphaevolve] — the AlphaEvolve kernel speedup cited in the AI-attribution register.
• [@yuksekgonul2026learning] — the test-time-training kernel result cited in the AI-attribution register.
• [@sherry2021fast] — the published base rate: about half of algorithm families show little or no improvement over decades, with improvements arriving at roughly 1.44 per family since 1940.