Skip to content
Language models

Qwen3-0.6B-Game-Decision-Model

a compact AI for decision-making across 23 games

3 min readupdated

A specialized language model from kaushik-harsh-99, fine-tuned to choose optimal moves across 23 board, card, strategy, and puzzle games. Built on the lightweight Qwen3-0.6B architecture (0.6B parameters), it merges LoRA adapters (small add-on layers on top of a base model) and an Unsloth FastDecisionModel / CLEF classification head into unified 16-bit weights. Decision accuracy jumped from 28.04% on the un-tuned base model to 67.81%, all in a package just over a gigabyte that runs even on low-end GPUs. It is built for developers of game bots, simulations, and agents who need fast action selection from text-based game state descriptions.

Key features

  • Trained on 1,092,754 game states from the boardgame-decision-models dataset spanning 23 games — from chess and poker to Minesweeper, 2048, and Mastermind.
  • +39.77% accuracy boost: averages 67.81% on a 2,193-sample benchmark, hitting near-perfect scores on individual titles — 100% in Mastermind, 98% in blackjack, and 96% in Yahtzee.
  • Ready-to-use merged weights: shipped as model.safetensors (1.14 GB + 52 MB classification head), so there is no need to manually merge adapters before running inference.
  • Low memory footprint: supports 4-bit quantization (weight compression to save memory) via Unsloth, fitting in under 2 GB of VRAM.

What you can do

  • Build a game bot or sparring partner: plug the model in via the ready-made GameDecisionRunner class to generate moves and score action probabilities in blackjack, Connect Four, 2048, or Dots and Boxes.
  • Add an action-selection module to an LLM agent: pass the current environment state and available choices to quickly get the best next move along with a confidence score.
  • Test it in the terminal: run the included test_inference.py script in interactive mode to test the model's logic on your own game scenarios.

Example

from runner import GameDecisionRunner

runner = GameDecisionRunner.from_pretrained("kaushik-harsh-99/Qwen3-0.6B-Game-Decision-Model")
result = runner.predict(
    game="blackjack",
    state="Dealer shows 7. Player hand: [10 of Clubs, 6 of Hearts] (Hard 16).",
    choices=["Hit", "Stand", "Double Down", "Surrender"],
    instructions="Select optimal blackjack basic strategy action."
)
print(result["best_action"], f"{result['confidence']:.1%}")
Read moreDetails
Benchmark (2,193 samples): overall accuracy is 67.81% vs. 28.04% for the base Qwen3-0.6B, with cross-entropy dropping from 1.5026 to 0.7298; the biggest jump is in Dots and Boxes (88% vs. 3%).
Hardware & memory: FP16 weights take 1.19 GB, and 4-bit mode uses under 2 GB VRAM; training took 70.3 hours on a single NVIDIA RTX 3050 (6 GB).
Training setup: 34,149 steps with the AdamW optimizer, 1,024-token context length, LoRA rank 16 and alpha 32.
Limitations: accuracy drops in complex strategy games requiring deep lookahead (chess — 29.2%, checkers — 35.0%, solitaire — 23.0%).
Examples
Discuss Telegram 2 44

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 1.1 GB · ready files1.8 GB
  • FP8 / INT8recommendedweights 0.6 GB1.2 GB
  • 4-bitweights 0.3 GB0.9 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

A ready script for the machine we advise above: paste it and the model downloads and starts.

  1. 1. Rent A30

    $0.26 an hour with the coupon. Pick the Ubuntu image when you create the machine.

    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.

    # kaushik-harsh-99/Qwen3-0.6B-Game-Decision-Model — A30, FP8 / INT8
    # NVIDIA driver: CUDA 13 needs 580+. If older or missing, the current nvidia-open is installed (keep yours: KEEP_DRIVER=1)
    DRV=$(nvidia-smi --query-gpu=driver_version --format=csv,noheader 2>/dev/null | head -1 | cut -d. -f1)
    case "$DRV" in ""|*[!0-9]*) DRV=0 ;; esac
    if [ "$DRV" -lt 580 ] && [ -z "${KEEP_DRIVER:-}" ]; then
      . /etc/os-release && \
      curl -fsSLO "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu${VERSION_ID//./}/x86_64/cuda-keyring_1.1-1_all.deb" && \
      sudo dpkg -i cuda-keyring_1.1-1_all.deb && rm -f cuda-keyring_1.1-1_all.deb && \
      sudo apt-get update && \
      sudo DEBIAN_FRONTEND=noninteractive apt-get install -y nvidia-open || \
      { echo "Could not install the NVIDIA driver: see the error above. To keep the current one, run with KEEP_DRIVER=1."; exit 1; }
      echo "NVIDIA driver updated. Reboot the machine (sudo reboot) and run this same script again."
      exit 0
    fi
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv venv ~/vllm --python 3.12 && source ~/vllm/bin/activate
    uv pip install vllm
    vllm serve kaushik-harsh-99/Qwen3-0.6B-Game-Decision-Model --max-model-len 8192 --quantization fp8
  3. 3. Use it

    The server listens on port 8000 and speaks the OpenAI API: http://localhost:8000/v1 on the machine. From your computer: ssh -L 8000:localhost:8000 user@machine-IP.

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.

  • MTP (multi-token prediction)0.34.0from November 16, 2026new release

    1.5–2× generation speed

    If the model has MTP layers (DeepSeek-V3/R1, Qwen3-Next, GLM, MiMo and others), it predicts several tokens per step itself.

    How to enable: vLLM / SGLang: --speculative-config '{"method": "mtp", "num_speculative_tokens": 1}'. Check the model card.

    Earlier generations (2)
    • EAGLE-3 — Speculative decoding with a small draft head. For models without their own MTP layers.
    • EAGLE-2 / Medusa — Earlier draft-head generations. Still work if that is all you find.
  • FP8 KV cache0.34.0from November 16, 2026new release

    2× more context in the same memory

    Keeps the attention cache in 8 bits, so long contexts fit where they did not.

    How to enable: vLLM: --kv-cache-dtype fp8; llama.cpp: -ctk q8_0 -ctv q8_0.

  • vLLM / SGLang0.34.0from November 16, 2026new release

    many times more tokens/s on a server

    Server engines: batching, FlashAttention, FP8 and speculative decoding out of the box.

    How to enable: vllm serve <model> — an OpenAI-compatible API.

    Earlier generations (1)
    • llama.cpp — GGUF models on any GPU, Mac or even CPU; some layers can stay in RAM.

Model family

  1. Qwen3-0.6B-Basefamily root
  2. Qwen3-0.6B
  3. Qwen3-0.6B-Game-Decision-Model

New posts about versions of Qwen3-0.6B-Base (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 3All on Hugging Face

About the model

kaushik-harsh-99/Qwen3-0.6B-Game-Decision-Model
Author
kaushik-harsh-99
Category
Language models
Released
Posted

On NeyroLearn

Hugging Face

442

Ask AI about this model

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

Comments