Skip to content
Other

Maia-3 5M LiteRT

human-like chess engine for mobile and web

2 min read

The model predicts the chess move most likely to be made by a real player with a given Elo rating. This is a ready-made conversion of the original Maia-3 5M checkpoint from the CSSLab laboratory at the University of Toronto into the LiteRT format (formerly TensorFlow Lite, a lightweight format for running AI models on mobile devices and in the browser). The model weights were not altered or fine-tuned: the file is about 20.8 MB and fully replicates the logic of the original PyTorch version. The solution is designed for cross-platform chess applications and suits game and trainer developers who need a realistic bot without a cloud server.

Key features

  • Full identity to the original: the divergence of legal move logits from the initial PyTorch model does not exceed 9.5e-06.
  • Competitor strength factoring: the model accepts the current Elo rating of the bot itself and its opponent as inputs, adjusting the playing style to the specified level.
  • Compact size: the resulting maia3-5m.tflite file takes up 20,813,796 bytes (~20 MB), allowing it to be embedded directly into applications.
  • Deterministic build: the conversion is performed via litert-torch 0.10.0 and is byte-reproducible when restarting the export script.

What you can do

  • Embed the file into mobile chess games on Android and iOS or into a web application for offline play against a realistic sparring partner.
  • Implement an adaptive trainer that adjusts to the user's rating (from beginner to experienced player) instead of playing at full strength like classical engines.

Example

The model is invoked via the serving_default signature with three input tensors: the board state args_0 of shape [1, 64, 97] (float32 — the last 8 positions with one-hot piece encoding), its own rating args_1 (int32 [1]), and the opponent's rating args_2 (int32 [1]). At the output, output_0 produces a logit vector of [1, 4352]: indices 0–4095 encode regular moves (from * 64 + to), and 4096–4351 encode pawn promotions to a queen, rook, bishop, or knight.

Read moreDetails
Base model: UofTCSSLab/Maia3-5M (5 million parameters, original file maia3-5m.pt)
Conversion tools: litert-torch 0.10.0, PyTorch 2.14.1, jax/jaxlib 0.11.2
Graph modifications: the torch.nn.RMSNorm layer is expanded into standard operations for compatibility with LiteRT
License: GNU Affero General Public License v3.0 (AGPL-3.0)
Discuss Telegram 0 11

Will it run?

Working out the memory needs

The bot is still collecting this model's weight sizes; the estimate will appear here in a few minutes. Meanwhile you can rent a GPU with headroom.

Massed ComputeOpen Massed Compute

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # ber4444/maia3-5m-litert
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download ber4444/maia3-5m-litert --local-dir ~/models/maia3-5m-litert
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to maia3-5m-litert
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of maia3-5m-litert (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

ber4444/maia3-5m-litert
Author
ber4444
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

110

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments