Historical material
Preserved research material.
Earlier formats, artifacts and limitations remain available; current entry points are in the main navigation.
Return to current downloads guide
Archived notebook content
Historical CM-14 checkpoint
The 86,966,016-parameter transformer reads clean, complete move history: one token per UCI move, including promotions. It has 12 layers, 12 heads, width 768, 512 context positions and 1,975 vocabulary entries. Move logits and fixed-White win/draw/loss logits are normalized separately. Outcome labels are not supplied as input, and the WDL diagnostic does not choose the browser move.
CM-14 began from fresh random initialization with a hybrid Muon/AdamW optimizer. This promoted checkpoint continues that CM-14 run; it is not a continuation of CM-13 weights.
Historical notebook material and artifacts
Current final snapshot
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.
Current input and serving interfaces · Final evaluation · Published browser package. Earlier explanations below retain their historical context; CM-10 is no longer the browser default.
Earlier notebook material and artifacts
Online on Lichess: CM-13, audited eight-hour snapshot. Play LittlePly on Lichess, with a different tokenizer and helper. CM-13 results and limitations.
Current online model: CM-13
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.
Eight-hour checkpoint, results and training context; serving metadata. The CM-10 explanations and artifacts below describe the separate browser model and earlier experiment.
Downloads / CM-10
Choose a checkpoint before downloading.
Current browser weights and evidence are versioned separately from historical training packages.
Browser model: CM-10
CM-10 checkpoint · Evaluation and limitations · Retired browser artifact · Reproduction instructions
Full curriculum payloads and resumable optimizer checkpoints are not bundled.
Earlier releases and downloads
Downloads / versioned evidence
Choose a checkpoint before downloading.
Match every artifact to its model and experiment. Availability differs across releases; old packages are not current-model bundles.
Historical model: CM-7
Start with CM-7 reproduction instructions for browser weights, source, hashes, training counts and all 120 evaluation PGNs. Full curriculum payloads and resumable optimizer checkpoints are not bundled. This is not a turnkey replay of training.
Checkpoint registry · Browse source · Replay or download individual games
Earlier releases
- CM-4 weights and training/evaluation artifacts.
- CM-6 evaluation PGNs and audited metrics.
- CM-5 is documented in EXP-7; it is not a complete downloadable training bundle.
CM-2 pilot and CM-3 baseline packages
The historical release below retains its original files and checksums.
No resume checkpoints, optimizer state or private human-test games. Large archives are split into numbered parts; join and verify them before extraction.
Reproduction instructions · Release manifest · SHA-256 checksums
Download and verify everything
curl --fail --output download.py https://artifacts.littleply.com/v1/2026-10-28/download.py python3 download.py
Inspect the download helper. It checks each part, joins split archives and checks the complete archive. Requires Python 3 and curl. Model export reader loads safetensors without optimizer state.
CM-3 / Monthly filtered raw PGN
430.49 MB compressed · 3 download parts
Part 1 of 3 (200.00 MB) · Part 2 of 3 (200.00 MB) · Part 3 of 3 (30.49 MB)
Complete archive SHA-256
870ac2af018b938d1c5de372fc7b7641927870d1114f702228c5d52016be74b9CM-3 / Monthly tokenized dataset
183.33 MB compressed · 1 download
Complete archive SHA-256
c29b80e37436266d3a7a371e4ee09c2009c474a460821c73322c55873ef7b320CM-2 / Pilot filtered raw PGN
0.45 MB compressed · 1 download
Complete archive SHA-256
5d657e97fc6ffdc8d19af3d58ab95c51b9c2012757584a37571f43f914b92ad8CM-2 / Pilot tokenized dataset
0.49 MB compressed · 1 download
Complete archive SHA-256
d678ffb1d80228477d2a4d80313bc401c5d58cb95369587450dae48021c95dcbCM-3 / Monthly model weights + config
319.75 MB compressed · 2 download parts
Part 1 of 2 (200.00 MB) · Part 2 of 2 (119.75 MB)
Complete archive SHA-256
f8f81e373cd2a68e2bbe9eb46148f2b75ca9f516f300efbef2fb735d1aa52913CM-2 / Pilot model weights + config
319.35 MB compressed · 2 download parts
Part 1 of 2 (200.00 MB) · Part 2 of 2 (119.35 MB)
Complete archive SHA-256
228fda53011822524fd5c513aaca809662e2639ab208dd9be07d8e744645319aCM-3 / Versioned source snapshots
0.05 MB compressed · 1 download
Complete archive SHA-256
f89a314aa665380c7017652e1504b509328d795f5d61d738450cc0c91e6248e6
What is reproducible here?
Exported tensors exactly match the original model tensors. Token/offset bytes and public corpus metadata are preserved; dataset manifest paths are sanitized with original hashes recorded. Pilot historical training source is unavailable. Historical launch receipts and machine configuration are not distributed, so this is not a turnkey or bit-identical cross-hardware training replay. Follow the linked instructions and source pipeline for the boundaries.
Licensing and full upstream data
Lichess database-derived data are CC0. Original project source is Apache-2.0. Read the licensing declaration and download the new Apache-2.0 source bundle with its manifest and checksums. Historical R2 archives are byte-for-byte unchanged; their older licensing notes are superseded for original code by this declaration. Model archives retain their existing recorded terms; this source-license addendum does not relabel data or third-party code. The 29.22 GB full September archive is not duplicated here: its official URL and SHA-256 are in the release manifest. The monthly filtered raw PGN contains the header-qualified staging input before final deduplication and context exclusions.