Ender38
Ignore the assets. Go for the weights file below.
This is a 64x6 nn trained to play 12 man endgames and below. Leela sucks on fen’s because they have no history. I use @dje’s no history PR to solve that problem. LeelaChessZero/lc0#305
Batch 38 (plus maybe 5 self-play batches) had best performance so far on my 149 position endgame test suite.
Ender 38 - 5 sec, 2 thr, no prune
Success rate: 62.42% (93/149)
And versus two networks head to head 1min+1sec, it was even with a late test10 256x20, and soundly beat a late test9 64x6. All 12 man positions are played both ways.
Score of Ender38 vs T11258: 33 - 33 - 42 [0.500] 108
Score of Ender38 vs T9149: 36 - 23 - 41 [0.565] 100