Skip to content

@amkozlov amkozlov released this Feb 22, 2021


  • check for square brackets in taxon names (not allowed by Newick)
  • fix 'LIBPLL-111: memory exhausted' error on very large ladder-like trees
  • fix parallel compilation of tests (#108)
  • do not print incorrect energy measurements after the latest Linux kernel patch
  • coarse-grained parallelization: fix bootstrap failure/hanging in some special cases (eg autoMRE{x} where x<50)
  • check for unifurcations in constraint tree (#114)
  • improve error message when invalid number of substitution rates is specified
Assets 6

@amkozlov amkozlov released this Sep 21, 2020


  • regression: RF computation with multiple files (--rf tree1,tree2) reported incorrect, unrealistically high distances (eg >0.8). However, --rf treeset.nw format was not affected and reported correct values.
  • load balancing failed in rare edge cases (<=1 sites/core)
Assets 6

@amkozlov amkozlov released this Aug 8, 2020

Major new features

  • coarse-grained parallelization across tree searches (--workers option) -> see details
  • adaptive parallelization (--threads auto{MAX}, --workers auto) -> see detais
  • print per-site log-likelihoods (--sitelh command, #26)
  • alignment column weighting (--site-weights option, #51) -> see details

Other features & optimizations

  • energy monitoring via RAPL -> see details
  • partial loading of RBA alignments (ExaML-style, saves memory for MPI runs on large supermatrices)
  • detect CPU core oversubscription to prevent extreme slowdowns -> see details
  • collapse short branches (bl=min) into multifurcations (#23)
  • write down intermediate results (mlTrees, bootstraps)
  • much faster (20x-30x) parsimony starting tree generation


  • asc bias failed under certain circumstances (site repeats disabled and non-DNA data)
  • fix zero subst. rates in FLU model (#93)
  • out-of-range branch lengths after scaling/normalization (eg #86)
  • base frequency optimization sporadically failed when estimates hit the boundaries


  • print CPU model/RAM/core count in logfile
  • if --blmin is specified, adjust --precision accordingly
  • remove BS support + 0.0 brlen at the root
  • ensure that min/max brlens are not exceeded in user starting tree
  • check that --msa file is not empty and not a directory
  • check if start trees are bifurcating (#96)
  • check if partition range has all sites in single-partition case
  • check if +BU{} is specified in single-partition case
  • do not use word "pattern" if pattern compression is off
Assets 6

@amkozlov amkozlov released this May 20, 2019

This is the first RAxML-NG release without "BETA" tag, which means we now consider it stable enough to be used in production analyses. We strongly recommend all RAxML8.x users to upgrade!

More good news: RAxML-NG application note has been accepted in Bioinformatics, so please update your citations!

New features

  • consensus tree computation (--consense command)
  • ancestral state reconstruction (--ancestral command)
  • custom DNA substitution models via rate symmetries (eg --model DNA010010 = --model HKY)


  • more accurate autodetection of CPU core number (in particular, for AWS instances)
  • improve file vs. number autodetection in --bs-trees option (#64)
  • add CMake install target (#66)
  • fix small memory leaks in libpll
Assets 5

@amkozlov amkozlov released this Mar 5, 2019

Optimizations and extensions

  • support model parameters in PAML format (--model PROTGTR{paml.txt}+G4)
  • TBE optimization (>100x faster thanks to @lutteropp)
  • safety checks can be disabled one-by-one, eg --force threads,.. or --force model_overfit
  • print trees on a (small) terrace into standard multi-line newick file
  • RF distance: accept multiple newick files separated by commas (eg. --rf tree1.nw,tree2.nw)
  • properly handle zero state frequencies via rate matrix reduction trick


  • segfault in load balancing (rare special case)
  • fix empirical p-inv computation with pattern compression
  • #60: a rare bug in Brent optimization with values close to 0 (reported by @ephofmann)
  • --loglh: user-specified brlen scaler values were ignored


  • use 20 random starting trees by default if topological constraint is specified
  • --evaluate: save best-scoring tree into bestTree file
  • MPI: broadcast random seed
Assets 5

@amkozlov amkozlov released this Jan 11, 2019

IMPORTANT: New defaults

  • 20 starting trees = 10 random + 10 parsimony (was: 1 random)
  • scaled/proportional branch length model for partitioned alignments (was: linked)
  • autoMRE bootstopping (was: 100 BS replicates)

New features

  • Robinson-Foulds distance computation (--rfdist / --rf)
  • print bootstrap replicate MSAs (--bsmsa)
  • flexible MSA character maps for multistate data (see here)
  • new AA substitution model: DEN (Kim 2018)


  • when using multiple starting trees, raxml.bestModel file used to store model params from the last tree search and not the one that yielded the best-scoring tree
  • deadlock in --loglh with multiple threads
  • fix CLV scaling with p-inv models (RAxML/ExaML seem to have this issue as well)
  • duplicate taxon name check was broken since v0.7.0 (reported by Terry Jones)


  • account for hyper-threading when detecting the number of CPU cores
  • improve error messages for FASTA/PHYLIP formatting issues
  • check if --prefix directory exists
  • disable overfitting check for single-gene alignments
Assets 5

@amkozlov amkozlov released this Oct 22, 2018

New features

Minor features and optimizations

  • IMPORTANT: site repeats are now enabled by default (in case of problems: add --tip-inner on & please report!)
  • improved load balancing: weight sites according to # states and # rate categories (in case of problems: add --extra lb-kassian & please report!)
  • thread pinning by default in hybid MPI/pthread mode (--extra thread-nopin to disable)
  • remove gap-only & duplicate sequences from the reduced alignment file
  • build parsimony trees in master rank only
  • check for invalid symbols (spaces,semicolons etc.) in taxon names (problems with Newick)


  • IMPORTANT: --support command was broken in v0.6.0, printing BS support values of 0
  • improve base frequency normalization in empirical AA models (#46, thanks to @bredelings )
  • increase precision of WAG rates (thanks to @bredelings )
  • check for terraces only in --brlen unlinked mode
  • fix model optimization with fixed branch lengths (--evaluate --opt-brances off)
  • fully resolved/bifurcating constraint tree was ignored (reported by Lois Rancilhac)
  • fix overflow in site repeats on very large unpartitioned alignments
Assets 5

@amkozlov amkozlov released this Jun 16, 2018

New features

  • support for binary (aka presence/absence) and multistate (aka morphological) models
  • comprehensive and partial topological constraints (--tree-constraint)
  • binary MSA format (a la ExaML)
  • phylogenetic terrace detection using terraphast library (thanks to @upsj et al.)
  • support for unlinked branch length mode in partitioned analyses (--brlen unlinked)
  • automatic MRE-based bootstrap convergence check (also a posteriori with --bsconverge)
  • outgroup rooting (--outgoup)

Minor features and optimizations

  • separate command to generate starting trees (--start)
  • (alias) command to compute tree loglh w/o optimization (--loglh)
  • console-only mode (--nofiles)
  • user-specified precision for logLH, branch length etc. (--precision)
  • --parse command to compress alignment patterns, generate binary MSA and estimate resource requirements (RAM, #threads)
  • random/parsimony/user starting trees can be combined in a single program invocation
  • parsimony starting trees on mixed-datatype alignments
  • faster restart from checkpoint by re-using existing starting trees and binary MSA file
  • add missing SSE3 PLF kernels (major speedup on old pre-AVX machines)
  • libll and pll-modules are now built with CMake (grand merci @BenoitMorel)


  • fix p-inv optimization in multi-threaded multi-partitioned runs (brent opt param broadcast)
  • fix scaled branch length optimization when branches hit min/max value
  • better error messages for ill-formatted partition files
Assets 5

@amkozlov amkozlov released this Nov 28, 2017

This release contains multiple important bugfixes and several minor improvements.


  • fixed ML estimation of the proportion of invariant sites (+I)
  • invalid bootstrap support values (>100) on datasets with 32N taxa (thanks to Shoyo Sato)
  • fixed a bug in multi-partition parameter optimization involving LG4X/FreeRate models
  • load balancing failed under very specific circumstances (thanks to Shoyo Sato)
  • FreeRate model optimization failed when rates/weights approached the min/max values (thanks to Fabricia Ferreira Do Nascimento)
  • fixed parallel reduction buffer overflow on alignments with 1000s of partitions
  • several memory leaks in model parameter optimization routines


  • new --check command: check alignment/partition file, remove all-gap columns (similar to -f c in RAxML8)
  • PHYLIP parser is now ~10x faster on very long alignments (>10M sites, thanks to Mozes Blom)
  • several improvements for short/small alignments (avoid full tree traversals, faster tip-tip lookup table pre-computation)
  • parallelization hints: issue a warning if too few alignment patterns per thread are assigned
  • add DNA model aliases to be compatible with ModelTest-NG
Assets 5

@amkozlov amkozlov released this Sep 10, 2017


  • bootstrap support values were consistently underestimated if and only if --all command was used with --bs-trees greater tha 250. Many thanks to Fabricia Nascimento for spotting and reporting this bug!

New features/enhancements:

  • --support command: bootstrap mapping aka computing BS support values using an existing set of bootstrap trees from a newick file (similar to -f b option in old RAxML)
  • site repeats optimization technique as described in (Kobert et al. 2016). This feature is still experimental and disabled by default; it can be enabled with --site-repeats on switch.
  • faster AVX2 kernels for protein data (5%-20% speedup)
  • automatically enable per-rate scalers on very large alignments (>2000 taxa) to prevent numerical underflow
Assets 5