Superseded descriptions preserved for reading old experiments.
Everything below is an archived snapshot. Words such as “current”, “live” and “default” refer to that snapshot’s date, not today. Return to the current CM-14 architecture.
CM-13 / final snapshot
From clean history to one complete move.
The model is shared; browser and online assistance differ.
Fresh random initialization with AdamW; clean game history, one token per complete UCI move (one ply), including promotions. The 86,966,016-parameter transformer has 12 layers, 12 heads and width 768, with a 512-token context and 1,975 vocabulary entries. No masking and no outcome labels in inputs. Move outputs and fixed-White WIN/DRAW/LOSS outputs are separately normalized from the shared transformer and tied output matrix.
Tokens and context
Input: BOS 1971 followed by complete move IDs 0-1967. Promotions have their own full UCI token. WDL IDs 1968/1969/1970 mean White win/draw/White loss. PAD 1972, unused QUERY 1973 and EOS 1974 are not outcome conditioning. The 512 positions include BOS; history is never silently cropped.
Policy and external assistance
The browser takes legal argmax over move logits and shows the separate WDL diagnostic. It has no search or engine assistance. Lichess uses a separately deployed Stockfish-assisted service; its helper settings can evolve independently of the model weights. The frozen EXP-17 evaluation used mate-only assistance, with outcome search off and without the later 97% endgame takeover. Do not interpret the benchmark as unassisted browser strength.
CM-13 starts from random initialization with AdamW. Clean game history is encoded as one token per complete UCI move (one ply), including promotions, in a 512-token context. There is no masking and no outcome label in the inputs.
A shared 12-layer, 768-wide, 12-head transformer and tied output matrix produce two separately normalized distributions: move outputs and fixed-White WIN/DRAW/LOSS outputs. Outcome predictions are diagnostics, not policy conditioning or a search value function.
The model selects ordinary legal moves. A separate asynchronous Stockfish 19 helper can intervene for explicit offensive or defensive mates; centipawn scores do not trigger changes. This deployed system is not search-free. Engine mate reports are not independently verified proof certificates.
Serving interface
Input is BOS 1971 followed by complete move-history IDs 0–1967. WDL IDs 1968/1969/1970 mean White win/draw/White loss. PAD 1972, unused QUERY 1973 and EOS 1974 are not outcome conditioning. Legal move ranking uses the move outputs with deterministic UCI tie-breaking. History is not silently cropped at the 512-token boundary.
The deployed model is the frozen eight-hour snapshot at update 148,481. Training was scheduled to continue independently until November 2, 2026 at 20:09:54 UTC; later training does not change the live checkpoint without a separate promotion.
Historical full-history architecture, still used by CM-10 browser play
Full-history model / CM-10
From game history to a legal move.
The learned policy and external mate search are distinct components.
The model
85,940,736 parameters; 12 transformer blocks, 12 attention heads, 768 embedding dimensions; 128 vocabulary IDs and 1,024 context positions. CM-10 uses the CM-5 square-triple codec: BOS 69 followed by source square, destination square and promotion-or-none for each ply. The full standard-start history is required. It does not accept CM-7 state-v2/FEN-only prompts.
The training MASK token 72 is not used in clean serving inputs. Full histories are retained; context exhaustion produces an explicit error rather than silently truncating the input.
The playing system
An authoritative chess board maintains legality and repetition state. The neural policy ranks legal moves by joint probability of their three tokens. A separate bounded mate helper may choose a proven mating move before querying the model. This search is external, not learned reasoning.
Lichess serves the approved inference export on CPU. Browser play uses FP32 WebGPU with a three-ply helper, bounded by 50,000 nodes and one second. Its helper configuration differs from the live evaluation, so browser strength is not established by the reported match results.
A token is one integer ID. The vocabulary size counts allocated embedding/output IDs, including reserved IDs. A context is an ordered sequence of token positions. A chess ply is one player moving; a full move normally contains two plies. Both families generate a ply as three tokens: source square, destination square, promotion.
Through CM-5: full game history
128 allocated IDs; 1,024 position capacity. Square IDs 0–63, promotion IDs 64–68 (64 means none), BOS 69, EOS 70 and PAD 71. IDs 72–127 are reserved: 72 defined IDs, not 128 distinct chess concepts.
Start with BOS and every played move from the standard start.
Encode each ply as three tokens; no silent history cropping.
Train dense causal next-token prediction on valid sequence targets; padding is excluded.
At inference, append source, destination and promotion conditionally.
The 1,024-position allocation is a token capacity, not 1,024 moves. BOS, ending tokens and candidate continuation tokens also need positions. Overflow is handled explicitly by the serving controller.
CM-6 and CM-7: structured state and bounded history
512 allocated IDs; 122 operating input slots. The backbone still allocates 1,024 learned position rows; that does not make the state input 1,024 tokens. The current codec fixes 120 prefix slots, up to 16 genuine recent plies and three move targets.
Build a prefix from current board/state and optional recent moves.
Keep absolute slot positions; left-pad missing history and mask those attention keys.
Feed 120 prefix + the first two target tokens = 122 model input slots.
Supervise only three logits, at zero-based positions 119, 120 and 121.
The logits predict source, destination and promotion respectively. Inference uses 120, then 121, then 122 input positions. There is no prefix-reconstruction loss and no EOS target.
Up to 16 plies × 3 tokens; missing slots are masked PAD.
65 slots
STATE marker + 64 square occupancy tokens, a1 through h8.
5 slots
Side, castling, en-passant, halfmove clock, NEXT marker.
3 targets
Source → destination → promotion. Only the first two are appended to the model input.
2 + 48 + 65 + 5 = 120 prefix slots; 123 serialized slots, 122 shifted input slots, three supervised targets. Fullmove number is metadata, not an input token.
What the 512 IDs mean
CM-6 and CM-7 retain IDs 0–71 (EOS 70 is defined but never emitted by this codec). IDs 72–127 remain reserved. Newly assigned IDs 128–378 cover square occupancy, side, castling masks, en-passant targets, halfmove-clock values and structural markers. IDs 379–511 remain reserved. Thus 323 IDs have assigned meanings, including legacy EOS; not every assigned ID is emitted in every valid position, and the raw decoder still produces 512 logits.
FEN supplies the current board/state, not an incomplete board position. It does not contain the whole move history or establish repetition history. The model can work without full past games; authoritative legality and repetition adjudication stay outside the model. Optional recent history does not reconstruct the complete past.
CM-6 introduced this input format; CM-7 retains it for curriculum and standalone-position training. Its objective and vocabulary differ, so its NLL must not be compared numerically with dense full-history NLL. Its 100-game evaluation also changed precision and helper policy: it is system-level evidence, not an input-only ablation.