jspaulsen/halluci-mate-v1a

TEXT GENERATIONConcurrency Cost:1Model Size:0.8BQuant:BF16Ctx Length:32kPublished:Apr 20, 2026License:mitArchitecture:Transformer Open Weights Cold

jspaulsen/halluci-mate-v1a is an alpha release 0.8 billion parameter chess LLM, based on the Qwen3 architecture, trained from scratch on the Lichess dataset. It utilizes a custom UCI move tokenizer and a 32,768 token context length. This model is specifically designed for generating chess moves, with a focus on constrained decoding to ensure legal move outputs.

Loading preview...

What is halluci-mate-v1a?

halluci-mate-v1a is an alpha release language model specifically designed for chess. Developed by jspaulsen, it is built upon the Qwen3 architecture with approximately 0.6 billion parameters and a substantial 32,768 token context window. Unlike general-purpose LLMs, this model is trained exclusively on the Lichess dataset, focusing on generating chess moves in UCI format.

Key Capabilities & Features

  • Chess-specific Training: Trained from scratch on human games from the Lichess database, reflecting human play patterns.
  • Custom UCI Tokenizer: Employs a unique tokenizer for UCI moves, including special tokens for game conditions (<WHITE>, <BLACK>, <DRAW>).
  • Constrained Decoding: Supports masking logits to legal moves, preventing the generation of illegal chess moves, which is recommended for practical use.
  • Qwen3 Architecture: Utilizes a Qwen3-0.6B base with 28 layers and 16 attention heads.

Limitations & Considerations

  • Alpha Quality: The model is in an early development stage; move strength and strategic robustness are unvalidated.
  • Illegal Move Potential: Without constrained decoding, the raw model may occasionally produce illegal move tokens.
  • Human Game Bias: Reflects idiosyncrasies and blunders present in human games, especially at lower ratings.
  • FEN Support: Limited support for analyzing positions from arbitrary FENs.