Kuna Decompiler vs IDA Pro, Ghidra and angr — Compared on DecBench

Every accuracy figure on this page comes from DecBench v1.0, generated on 2026-09-19, which runs 14 decompilers over 96,092 functions from 770 binaries built from 40open-source projects — bash, coreutils, u-boot, OpenSSH and the rest. Nothing here re-runs that benchmark and nothing here invents a score. The counts were read from the scoreboard's own data payload with its “normalize failures” toggle off, which is the site default.

DecBench publishes three metrics per corpus. Structureis a graph edit distance between the decompiler's control-flow graph and the ground truth: the count of node and edge insertions, deletions and substitutions needed to turn one into the other, scored as the fraction of functions where that distance is zero. Typesis the fraction of variables and arguments whose recovered type matches the binary's DWARF information. Recompile is the fraction whose decompiled C rebuilds to byte-identical assembly. All three are reported as hits over the functions the metric could be evaluated on, which is why the denominators differ between columns.

The tool runs two stages. Stage one eliminates. A decompiler is removed, with the blocker named, when its published architecture list does not cover your target, when it does not run on your operating system, when its licence forbids work you are paid for, when its price exceeds your ceiling, or when it lacks a capability you marked must-have. Nothing is scored down for these — a licence that forbids your work is not a small deduction.

Stage two scores the survivors as score = 50·B + 30·F + 20·C. B is benchmark fit: the weighted sum of the three metrics under your chosen priority, then divided by the best surviving raw value so the leader scores 1.0. F is feature fit: the share of the capability list, excluding your must-haves, that the tool documents. C is cost fit: 1 − price/ceiling against a finite budget, 1 − price/highest-price without one, and 1.0 when everything surviving is free. The three weighted terms are printed beside every result so the ranking can be audited rather than taken on trust.

The 50/30/20 split is a stated editorial judgement, not a derived constant. Measured accuracy carries the most weight because it is the only term an independent third party produced; cost carries the least because it is the term you can most easily overrule with context this page does not have. A tool with no DecBench row is never given an invented B. It is scored on features and cost alone and listed in a separate group, because a tool nobody has measured is not the equal of one measured poorly.

Rupee figures are presentation only. They are the vendor's USD list price multiplied by the rate in the calculator, and they never enter the score — so a ranking does not quietly change when the exchange rate moves.

The release post, published on 29 July 2026 by Zion Leonahenahe Basque, quoted 44.4% perfect structuring against IDA Pro 9.2's 45.7% — Kuna 1.3 points behind. The DecBench v1.0 scoreboard generated on 2026-09-19has Kuna 1.515 at 48.4% against the same IDA build's 46.0%, so the sign of the gap has flipped in under two months. That is the point of the project rather than a footnote: it is refined against the benchmark continuously, so any figure quoted for it carries a date.

The author's own caveat still holds. Structuring is one dimension, and types, optimisations, recompilability and variable identification are listed as open. The calculator's type and recompile columns show what that costs. The release post published no licence; the repository does, and it is Apache-2.0.

One more thing worth knowing before you read any decompiler leaderboard: DecBench's 250-function sample set is public, and the benchmark itself warns that an LLM's score on it can reflect recall rather than decompilation. The coding agents top that set and are nowhere on the full corpus, so they are kept out of the ranking above.

• DecBench — scoreboard, metric definitions and corpora (v1.0, 2026-09-19)
• Kuna release post — Zion Leonahenahe Basque, 29 July 2026
• Kuna repository — licence, interfaces, vendored SLEIGH processor specs
• Hex-Rays — IDA pricing and tier comparison
• Hex-Rays — IDA Free non-commercial terms
• Hex-Rays — IDA Home non-commercial terms
• Vector 35 — Binary Ninja editions, prices and licence restrictions
• NSA — Ghidra, Apache-2.0, processor modules
• angr — BSD-2-Clause; architectures via archinfo, loaders via CLE
• r2dec — architecture modules and licences
• Fraunhofer FKIE — dewolf, a Binary Ninja plugin
• Avast — RetDec, MIT, supported formats and architectures
• Reko — supported architectures and formats

Consistency check:every one of the transcribed benchmark rows is re-derived on each page load against DecBench's own definition of its headline column — a function counted as perfect on at least one metric must also be counted on one of the three individual metrics, and cannot be counted more times than the three added together. All rows pass, on all four corpora.

Every benchmark count, price, licence and architecture list on this page was transcribed from the sources above on 2026-09-21. Figures a source does not publish are shown as not documented and are never estimated. This page does not host, mirror or link to redistributed builds of any commercial tool, and it gives no view on whether reverse-engineering a particular target is lawful where you are — read the vendor's own terms.

Background reading: an LLM wrote a decompiler that nearly matches IDA Pro — why the benchmark, not the model, is the interesting part.