86M-param decoder-only transformer plays chess on Lichess, full notebook public
andreisavu · x · 2026-10-29
Andrei Savu published a research notebook for LittlePly, an 85.9M-parameter decoder-only transformer chess model trained on a local Lambda Vector Workstation (with his OpenAI dot, Glimmer). It ingests full game histories and outputs next-move probabilities using a 128-token move codec (from/to/promotion), 12 decoder blocks, and a 1024-token context, with no reasoning trace; legal filtering and a disclosed three-ply mate helper complete the playing system. The bot plays on Lichess (up to 4 simultaneous games, bullet to classical), and models, datasets, and source are downloadable.
More from coding & agent
- Harbor v0.25 ships anti-reward-hacking toolkit: tampering detection, cheating prevention for agent evals — ajratner · 2026-10-29
- Anthropic says no regression behind Claude Code usage complaints, blames heavy sessions — bcherny · 2026-10-29
- LlamaIndex launches Document Parsing Protocol: 140 rules to standardize OCR output for agents — llama_index · 2026-10-29
- exe.dev Ships WIF Support: Agents Make OAuth Token Exchange Finally Usable for All — davidcrawshaw · 2026-10-29
- Protecting action-taking agents from prompt injection: move defense out of the model layer — Informal-Dust4499 · 2026-10-29
- Designing agent approval screens: show the effect, not the narrative — ClickOk5811 · 2026-10-29