It reached 4d on kgs server (account Pachi2 cluster version) and also runs well on lightweight hardware (raspberry pi, 2d). dcnn support doesn't require a gpu.
See latest go software updates for the latest update news.
The source code, written in C, is available on github.
Windows binaries are available in the release section.
The successive versions of the program are named after famous Japanese players:
Pachi 2.00 September 2008 Sana? Pachi 3.00 October 2009 Sanetsu Pachi 4.00 October 2009 Doetsu Pachi 5.00 Februari 2010 Dosaku .... Pachi 8.00 May 2011 Shuhaku Pachi 9.00 October 2011 Hakugen Pachi 10.00 August 2012 Satsugen Pachi 11.00 April 2015 Retsugen Pachi 12.00 January 2018 Jowa
How to use Pachi on Windows: See readme in windows release.
(instructions in Russian for older release): http://10kyu.ru/posts/go-bot-pachi-dlya-windows/
You can also check out the derivatives:
In Reinforcement Learning with Tensor Flow by Sayon Dutta (ISBN: 9781788835725), Pachi is incorrectly described as the strongest AI in the pre AlphaGo era "The AI programs of Go before AlphaGo totally relied on Monte Carlo Tree Search. These were Fuego, Pachi, Zen, and Crazy Stone. Out of which, Pachi was the strongest until AlphaGo defeated it using only policy network without using any search methods."