CM-4 supplementary evidence and downloads. EXP-5 training record · EXP-6 matched evaluation.
Later update: the Elite continuation and matched evaluation are complete. This foundation report preserves the earlier run and its reproduction artifacts.
Research update / 29 October 2026
A broader foundation. A controlled comparison.
A fresh four-hour run on six months of Lichess games, evaluated against the CM-3 checkpoint.
The learned model remains next-move imitation with no reasoning traces. These evaluation games used the same three-ply helper for both checkpoints. They do not evaluate the later Rust helper.
Read → inspect → reproduce
- Training record: data, configuration, exposure counts and checkpoint lineage
- Matched validation and frozen Sunfish comparison
- Replay and download all 200 games, including the unfinished game
- Training and evaluation source snapshot
- Sanitized training evidence, configuration and comparison evidence
Which weights?
The best foundation checkpoint is 857bc93bfffb287dc67e2845a14c214a4691bc1a008dae1fa4d044e5fccc87d2 at step 72,198, after 526,834,590 prediction targets. Its expanded-validation NLL is 0.573903. The final checkpoint at step 75,558 consumed more targets but did not complete a final validation; these are different weights.
Browser play still runs the CM-3 checkpoint 6133648b2926 and its original three-ply helper. It has not been replaced by this foundation model.
Reproduction boundaries
The source snapshot records original and rendered hashes; two private fallback dependency paths are redacted. Install dependencies in your own environment. The subset is inspectable source, not a turnkey training distribution. Existing model and dataset downloads remain the September release. The foundation model-only export and source packages are linked below. Six-month bulk data are not included; do not use the historical September dataset as if it reproduced this run.
Download the foundation release
Reproduction instructions · Versioned manifest · Checksums
Model-only FP32 archive: 319.75 MB compressed. Download part 000 (200 MB) and part 001 (119.75 MB), concatenate in order, verify, then extract. No optimizer or resume state. Complete archive SHA-256: ef717d2c25ef276fb3f160a0d382384ad2d24ae351d101a2a0fab2d08eba6c95.
Training, evaluation and live-service source subset · Complete Rust helper corresponding source with vendored dependencies.
Original project training/service code retains Apache-2.0. The Rust helper component is GPL-3.0-or-later, with dependency-specific notices. The corresponding-source archive records one test-only path redaction; Rust and vendor sources are unchanged. Read helper notices.
Next: Elite continuation
The proposed second stage continues from a broad foundation on Nikonoel Elite, as supplied: June 2020–November 2025, 66 months, without additional BOT/rating/time-control filters. Acquisition is reported complete (7.14 GB); preparation is now focused on a bounded newest-first subset. No token counts or trained continuation results are available in this report. Stronger tactical or strategic behavior is a hypothesis, not demonstrated reasoning ability.