Repository navigation
Request for Python API and ponder time examples #6
Description
First of all amazing work on this, I've been a big fan of Maia1/2 and playing against 3 online feels much more human!
Firstly, I was wondering if you would be able to provide some examples on using maia3 as a python package similar to maia2? The UCI interface isn't sufficient for what I'm trying to do, and am trying to reverse engineer things but probably a nicer way!
Secondly, from what I can tell, you are actively modelling ponder/move times (IE how long a human would take to make a move)? The ponder head doesn't seem to be giving me very good results, hoping you can advise.
I notice include_time_info is always set to false in the configs, do the released models support the time feature during inference?
Great work as always!
Reacted by Marco
Activity
daniel-monroe commentedon May 31, 2026
Hi Torbet,
The models on Huggingface all were trained with
--include_time_info=falseand do not support ponder time prediction. I have pushed changes at https://github.com/daniel-monroe/maia3/tree/rmponder that set the ponder value to 0 for base models, in addition to themaia3-23m-pondermodel which does accepts time control information and outputs ponder predictions and is currently used by the maia website. We are discussing a package at the moment and will continue to work on this.Torbet commentedon May 31, 2026
Amazing, thank you @daniel-monroe! I believe UCI supports an
infodict in the results, this may be a good place to include the ponder information if you wish to support it via the existing API.Will play around with the 23m ponder model! Thanks again :D