What Changed

Sequencing technology · AI · Benchmarking · Genomics · ComparisonModels of May 2023 to May 2026 · record to September 2026

• 0DNA SUP models in the v6.0.0 generation — RNA has one
• ~43%fewer read errors for sup@v5.2.0 than hac@v6.0.0, on five bacterial genomes
• 13/13isolates with dnd or dpd modification systems whose HAC v6 assemblies went wrong, in a 294-isolate study
• ~1 Qby which sup@v5.0.0 overstated its own quality scores, according to Oxford Nanopore
• 2.7–5.7×how much longer SUP took than HAC on the same data, in the same run

What changed, and how firmly each change is established standing first; rows are grouped by standing and never interleaved

In What Order

The order it happened in releases are dated by Dorado’s own release notes and measurements by publication; the v3.x models, for the older R9.4.1 pore, sit outside every comparison here and are not dated

1. 18 May 2023Dorado 0.3.0 ships the v4.2.0 models for 5 kHz data.
2. 24 Oct 2023Ryan Wick tests the v4.2.0 fast, HAC and SUP models and a research model fine-tuned for native bacterial DNA, on nine bacterial genomes. The research model edges out SUP.
3. 5 Dec 2023Dorado 0.5.0 ships v4.3. On 18 December Wick measures SUP at a median Q20.5, up from Q18.5, and assembly errors across the nine genomes down from 173 to 37; he reports that the bacterial training of the research model has been folded in.
4. 21 May 2024Dorado 0.7.0 ships v5: SUP moves from LSTM to a transformer, HAC stays LSTM. At London Calling, Oxford Nanopore puts v5 SUP at Q26, 99.75%.
5. 21 Jan 2025Dorado 0.9.1 speeds up basecalling. One laboratory’s SUP v5.0 rebasecall of a single run drops from 46.3 hours to 18.7.
6. 21 May 2025Dorado 1.0.0 ships v5.2: HAC grows from five LSTM layers to seven; SUP keeps its architecture and changes its weights. The R9.4.1, 4 kHz and RNA002 models and Fast5 input are dropped.
7. 27 May 2025Wick: HAC has about 15% fewer read errors and 65% fewer median assembly errors; SUP reads are essentially unchanged.
8. 13 Jun 2025Prior and colleagues, in the Journal of Clinical Microbiology: on 80 isolates, SUP v5.0 brings cgMLST typing close to Illumina, where v4.3 was sometimes insufficient.
9. 27 Jan – 26 Feb 2026A user reports on Dorado’s issue tracker that the same reads score lower under sup@v5.2.0. On 16 February Oxford Nanopore answers that v5.0.0 overstated its scores by about one Q point.
10. 18 Mar 2026Oxford Nanopore’s blog explains the transformer SUP and says it kept the LSTM’s inference-compute budget.
11. 24 Mar 2026Biggel and colleagues post a preprint testing SUP v5.2 on 294 foodborne-pathogen isolates.
12. 20 May 2026Dorado 2.0.0 ships v6.0: for DNA, a HAC model with custom recurrent layers and no SUP. Mike Vella, Oxford Nanopore’s VP of Machine Learning, at London Calling: “Importantly, we’re not releasing a sup model this time… we believe that this particular model is good enough that a sup tier is no longer needed.”
13. 11 Jun 2026Wick: hac@v6.0.0 has about 19% fewer read errors than hac@v5.2.0, but sup@v5.2.0 is still about 43% ahead; one Klebsiella genome produces about 100 errors per assembly with v6.
14. 19 Jun 2026Wick polishes the assemblies: the gap narrows and does not close. Vella suggests to him that phosphorothioate backbone modifications, absent from training, cause the Klebsiella errors.
15. 6 Jul 2026The preprint’s second version adds HAC v6: every one of 13 isolates carrying dnd or dpd DNA-modification systems gives a highly inaccurate assembly.
16. 8–9 Jul 2026Wick’s methylation test finds no model best on every measure. The next day he attributes the Klebsiella errors to 7-deazaguanine modifications from a dpd system, citing Biggel’s preprint.
17. 13 Jul 2026Dorado 2.1.0 speeds up v6.0.0 HAC on Blackwell GPUs, and v5.0.0-onward SUP on short reads.
18. 15 Jul 2026Kulkarni and colleagues, in Nature Communications: the older v4 model remains the reliable choice for CpG methylation, while newer models are better for other modifications.

The Argument

Two cases about the same model Oxford Nanopore’s for HAC v6, and Ryan Wick’s against dropping SUP; both are published, and both are named here

• Case 1

  Oxford Nanopore: HAC v6 makes SUP unnecessary

  • Custom recurrent layers keep the model close to the old HAC’s size: 9,808,240 parameters against 8,790,768.
  • Claimed: polished bacterial assemblies comparable to SUP v5.2, and assemblies above Q60 after polishing.
  • Claimed: better short-tandem-repeat calls, and the most accurate modified-base calling Oxford Nanopore has produced.

• Case 2

  Wick: on bacteria, SUP is still the better model

  • Median read accuracy Q20.6 for sup@v5.2.0 against Q18.1 for hac@v6.0.0.
  • Median errors per assembly 4 against 11, and polishing does not close the gap.
  • Methylation: no model best on every measure; for some 5mC work hac@v5.2.0 may be the right choice.

• Common ground

  What neither side disputes

  • hac@v6.0.0 is the most accurate HAC model yet, and in Wick’s run slightly faster than hac@v5.2.0.
  • SUP costs several times HAC’s compute; HAC is the tier that keeps up with a fully loaded PromethION.
  • Whether a DNA SUP v6 will ever ship is not stated by either.

What Others Add

The measurements, side by side each row is one post’s own comparison; posts ran on different Dorado versions and depths, so rows are not chained

• Quality scores

  A lower score is not a worse read

  the same 27,751 reads, basecalled twice

  • Under sup@v5.2.0, 62.5% fewer of them passed Q20 than under sup@v5.0.0, and 17.8% fewer passed Q15; lengths and base counts barely moved.
  • Oxford Nanopore’s answer: v5.0.0 scores ran about one Q point above the reads’ true accuracy, and v5.2 fixed the calibration.
  • Measured by alignment, SUP read accuracy did not change between the two, which fits that answer.

• Atypical modifications

  What HAC v6 was not trained on

  294 isolates, ten foodborne pathogens

  • Isolates carrying dnd (phosphorothioation) or dpd (7-deazaguanine) systems: all 13 went wrong with HAC v6, with 24 to 314 mismatching alleles, median 78.
  • SUP v5.2 is not immune: it failed on four Salmonella Kentucky isolates with a dnd system, at 20 to 68 mismatching alleles.
  • In public genomes, dnd turns up in 44 of 67 pathogen species and dpd in up to 8% of several Klebsiella, Escherichia and Enterobacter species.

• Methylation

  Modified bases do not follow the version order

  bacterial, plant, mouse and human data

  • For CpG methylation the older v4 5mCG model is the most reliable; the v5.2 5mCG model misses methylated CpGs surrounded by other CG residues.
  • Newer v5 models are clearly better for 5mC outside CpG, for 4mC and for 6mA.
  • For CpG, HAC performs almost exactly as SUP does; for 6mA and 4mC, SUP is consistently better.

Was It Good

Which model for which need a comparison of models ends in fit, not in a winner: each card says who it suits, what it is best at, and what it costs

• sup@v5.2.0

  For bacterial assembly and typing, where the GPU allows

  • The best result on every independent bacterial test here: reads, assemblies and polished assemblies.
  • Near-identical cgMLST profiles for 97.3% of 294 foodborne-pathogen isolates at 50× coverage.
  • The cost: 2.7 to 5.7 times HAC’s basecalling time — and it still fails on some genomes carrying dnd systems.

• hac@v6.0.0

  For high-volume, compute-bound and real-time work

  • About 19% fewer read errors than hac@v5.2.0, at slightly higher speed.
  • The tier Oxford Nanopore recommends for most users, and the one that keeps pace with a fully loaded PromethION.
  • The cost: it fails on genomes carrying dnd or dpd systems — 13 of 13 in one study. A reference-free check such as alpaqa flags affected assemblies.

• v4 5mCG model

  For CpG methylation

  • The most reliable model for CpG sites in an independent benchmark across bacteria, plants, mouse and human.
  • HAC performs almost exactly as SUP does for CpG, so the cheaper tier suffices.
  • The cost: a modification model is paired with its own basecalling model, so this choice ties the CpG calls to a v4 basecall.

• v5 modification models

  For 6mA, 4mC and 5mC outside CpG

  • Clearly better than the v4 models for these marks.
  • SUP is consistently better than HAC for 6mA and 4mC.
  • The cost: on bacteria no v5.2 or v6 model was best on every measure, and for some 5mC work hac@v5.2.0 had the best recall.

• Re-examine any quality-score filter after changing model. On identical reads, sup@v5.2.0 put 62.5% fewer over Q20 than sup@v5.0.0 — a recalibration, not lost accuracy. Check aligned accuracy, not the score.
• Test a new model on your own data before switching. One SUP release brought no read-level gain, and the newest HAC fails on a whole class of genomes.
• Polish bacterial assemblies with the bacterial model. In Wick’s tests the older pileup bacterial model made 6 of 147 assemblies worse, against 22 and 14 for the two read-level models, and improved the most.
• Compare studies within one pore. Models are trained for one sequencing condition each, so R9.4.1 data needs an R9.4.1 model — and Dorado 0.9.6 is the last release that carries them.