Understand what you are playing
A neural model predicts moves. Chess rules and a separate mate-search helper turn those predictions into play.
Take the short tourLittlePly / open chess research
A small chess model. An open notebook.
LittlePly asks how much chess a small neural model can learn from examples. Play against it, see how it works, and follow the experiments behind each version.
LittlePly is an experiment in AI research guided by Andrei Savu, working with his OpenAI dot, Glimmer, on a local Lambda Vector Workstation.
Play CM-7 on Lichess Try CM-7 in your browser
Live: CM-7, a model with 86.24 million learned parameters that predicts a move from the current board and recent history with a separate bounded mate helper. It produces no reasoning trace. See the architecture and move flow.
Before you challenge the bot
Casual or rated standard-start chess, bullet through classical, including zero increment. Up to four simultaneous games; bullet runs alone. No UltraBullet, correspondence or variants. Availability depends on the local host. The Lichess link opens an editable challenge dialog; sign-in and confirmation may be needed.
Why CM-7? Current board/state plus optional recent history supports standalone-position and endgame training. We adopt that input structure, without claiming clear superiority over CM-5. Final evaluation: 21W / 19D / 58L / 2 unfinished. Read the curriculum and selection evidence.
First visit
Start with a game. Follow the evidence.
See what changed
Read the question, training setup and full results behind the current CM-7 checkpoint, including losses.
Read the current evaluationInspect or reproduce
Trace the implementation to versioned source, downloadable weights and explicit gaps in the available artifacts.
Start with CM-7 reproductionState-input systems / adaptive 1/3/5/7 helper
The next research branch.
The same 100-opening/color schedule and adaptive helper; a system comparison, not an isolated model or theory-stage ablation.
21W / 19D / 58L / 2 unfinished
13W / 7D / 80L / 0 unfinished
CM-7 adoption favors state-based training flexibility, not a demonstrated win over CM-5. The historical full-history group below used a different three-ply helper. Never merge the groups into a model-only trend.
One frozen benchmark · three checkpoints
A measured research trajectory.
100 games per checkpoint against the same Sunfish version, openings, budgets and three-ply helper. These are observed results on a small selected suite, not Elo or proof of general playing strength.
WinsDrawsLossesUnfinished
27W / 9D / 63L / 1 unfinished
Curated-data subset · 3.72M training games
5h 58m · 940.65M additional total target exposures
21W / 9D / 69L / 1 unfinished
Six-month corpus · 4.14M training games
3h 58m · 551.41M total target exposures
13W / 2D / 85L
One-month corpus · 684,367 training games
86.01 minutes · 305.19M total target exposures
Total exposures describe completed jobs; validation-selected checkpoints are earlier and are identified in each report. Unfinished games are not draws. CM-5’s +5–7 points over CM-4 are outcome bounds, not confidence intervals; paired uncertainty includes zero. The earlier 20-game run used a different suite and is excluded here. Production’s iterative seven-ply helper is not this benchmark’s three-ply helper.
The work behind CM-7
Read newest first, or follow the training story from the initial curriculum through the follow-up to evaluation. A CM number identifies saved model weights; an EXP number identifies an experiment, not a newer model.