EXP-12 / completed
CM-7 / live Lichess model
State-conditioned curriculum model
Continued from CM-6 through a finite mixed curriculum and opening/endgame/Elite follow-up; selected by validation NLL at follow-up update 2,232.
86,235,648 parameters. 120-slot prefix encoding current board/state plus optional 16-ply history; up to 122 inference positions, with 512 allocated vocabulary IDs. Complete past game history is not required, and repetition history may be unknown. Different training objective; loss is not comparable to full-history runs. No reasoning traces.
Checkpoint SHA-256 29e68a8e85336cbb1f56efdad36bd76217728b6a376d1e5f74e89887d21383da
Initialized from CM-6.
Experiments
EXP-11 / completed
Opening and exact-endgame follow-up
Read experiment →EXP-10 / completed
Finite state-conditioned curriculum
Read experiment →Inspect and reproduce
Recorded evidence and source · All versioned source, including Rust · Published downloads and limitations · Games and checkpoint filters
Selected by validation NLL at follow-up update 2,232. Adoption is motivated by standalone-position and endgame training support, not demonstrated superiority over CM-5. The final system evaluation is 21W / 19D / 58L / 2 unfinished.