wrong repository

WARNING

This is not the LeelaChessZero repository, but it's older predecessor leela-chess. Go here for the LeelaChessZero releases.

v0.10

(This is identical to v0.10 RC1, you don't need to re-download.)

Note: There may be up to 24 hours where the server displays you are using lczero engine version 9 even after you upgrade to 10. This will be fixed soon(tm).

Main change:

• #577 - Two inputs to the Neural Network (rule50 and the all 1s plane) were not connected properly.

We don't expect this bug fix to resolve the current strength fluctuations, but it's still an important fix so we made another quick release. Thanks to Alexander Lyashuk and Trevor for both finding this issue independently, and Tilps for implementing a fix.

lczero10_mlk_bmi2_and_popcnt.zip is the Intel MKL release. Use this is you do not have a GPU and have a modern Intel processor (Haswell or later -- it will just crash if it's not new enough). You must download the dlls from Intel: https://software.intel.com/en-us/mkl

Thanks to the contributors to this release:
Alexander Lyashuk
Andy Olsen
Gary Linscott
Tilps

v0.9

(This is identical to v0.9 RC1, you don't need to re-download.)

Main changes:

• #546 NNCache full history
  • More accurate results (important for low playouts i.e. self-play games).
  • NPS for match games will be lower due to this, but more steady.
• #536 Disable tree reuse in training
  • More accurate application of Dirichlet noise to help discover new moves, for very small performance hit
  • NPS for client self-play will be much higher, but wall clock time to play a game slightly longer.
• #529, #526 Fix resign code
• #521 Fix --playouts
• #518 increment tbhits on root probes
• #516 faster loading of weights
• #512 rename visits to nodes (--visits is deprecated, use --nodes or go nodes N) instead

lczero_v0.9_avx2_mkl.zip is the Intel MKL release. Use this is you do not have a GPU and have a modern Intel processor (Haswell or later -- it will just crash if it's not new enough). You must download the dlls from Intel: https://software.intel.com/en-us/mkl

Thanks to the contributors to this release:
Alexander Lyashuk
Andy Olsen
Ankan
David Powell
Error323
Ganesh Krishnan
Gary Linscott
Gian-Carlo Pascutto
Tilps
dubslow
evalon32
jjoshua2

v0.8

(This is identical to v0.8 RC1, you don't need to re-download.)

v0.8 main changes:

• Fix a bug in First Play Urgency (FPU)
  • Do not include Virtual Loss (VL)
  • Also set FPU reduction to 0.1
• Fix a bug in 50 move rule for official test matches
  • Draw on move 50, not on ply 50.
• Add support for resigning self-play games
  • Server will enable this soon after v0.8 is enforced
  • See data collected regarding this here
• Add support for syzygy endgame tablebase
  • --syzygypath (UCI option will be in v0.9)
• Tune PUCT parameter
• go movetime N works now, except it moves too fast

Thanks to the contributors to this release:
Alexander Lyashuk
Andy Liu
Andy Olsen
Ankan
Ankan Banerjee
Brandon Lin
David H. Flint
Dubslow
Emil Kattainen
Francis Li
Gary Linscott
Gian-Carlo Pascutto
Joseph Ellis
Michael Pang
Robin Houston
TFiFiE
Tilps
Victor Popovici
andyliucode
evalon32
jjoshua2
jkiliani

v0.7

v0.7 has a critical bugfix to the Network evaluations affecting both training and gameplay. Please upgrade!

It also has improvements to our UCI compatibility, time management, and evaluation scores.

Thanks to the contributors to this release:
Akababa
Alexander Lyashuk
Andy Olsen
Gary Linscott
Joseph Ellis
Manuel Pégourié-Gonnard
MTGOStark
Preston T
Tilps
Victor Popovici
syjytg

v0.6

v0.6 is mostly bug fixes, fixing some crashes. In addition:

• lczero can read compressed network weights.
• Changed the cp eval function.
• Updated version of libopenblas.dll, which runs faster when using a large number of threads. This required including a few more supporting dlls.

Contributors this release were (in alphabetical order):
Andy Olsen
Arseny Krasutsky
Error323
Gary Linscott
Tilps

v0.5

v0.5 brings important changes to training data generation for speed, and critical bugfixes in the way that history was sent to the net and moves were received. We also now have fractional temperature support for variety in match games, so we should see more opening variety.

Upgrading is required - however, thanks to some great work by Andy Olsen (killerducky), the lczero client remains backwards compatible with old network files.

For UCI users, the engine now defaults to "weights.txt" if no weights parameter is supplied. Also, you can use the "--uci" flag to have it delay initializing until isready is sent, which should fix problems loading in GUIs like Fritz. Time management is also improved.

Contributors this release were (in alphabetical order):
Akababa
Alexander Lyashuk
Andy Olsen
Björgvin Ragnarsson
Brandon Lin
Douady Pâris
Error323
Gary Linscott
Shail Choksi
Tilps
jjoshua2
jkiliani
vchen30
xoto10

v0.4

v0.4 contains some client improvements to be more robust, and faster uploading games.

It also has tree re-use now, which should provide a substantial speed boost to game generation, and gameplay.

Contributors this release were (in alphabetical order):
Andy Olsen
Douady Pâris
Error323
Gary Linscott
davidsoncolin
davidxu90
evalon32
gsobala
jjoshua2

v0.2

v0.2 includes match support in the client, and a ton of upgrades to the lczero executable, including Neural Network caching, and time management support for UCI mode!

Windows GPU build is included here as well.

(NOTE: Updated to include bugfix for match games getting stuck on draws -- client is now v3).

v0.1 client

Initial binary release of client.