Godoter3.8-27B

16-bit merged weights and GGUF quantizations. A QLoRA fine-tune of Qwen3.8-27B on the Godot Engine 4.x documentation. Trained to answer questions about the Godot game engine with deep, accurate knowledge of the engine's APIs, nodes, and workflows.

Run with llama.cpp / llama-server.

Raw safetensors weights - GGUF quantizations

Quantizations

All quantizations ship with the vision projector (mmproj) required for image understanding - the mmproj file must be loaded alongside the main model.

Usage with llama-server

llama-server \
  --model /path/to/Godoter3.8-27B.Q4_K_M.gguf \
  --mmproj /path/to/Godoter3.8-27B-mmproj.gguf \
  --host 0.0.0.0 --port 8090 \
  --ctx-size 32768 \
  --n-gpu-layers 999

For Q8_0 on dual GPUs:

llama-server \
  --model /path/to/Godoter3.8-27B.Q8_0.gguf \
  --mmproj /path/to/Godoter3.8-27B-mmproj.gguf \
  --host 0.0.0.0 --port 8090 \
  --ctx-size 32768 \
  --n-gpu-layers 999 \
  --split-mode tensor \
  --tensor-split 1,1 \
  --main-gpu 0

Then query it with any OpenAI-compatible HTTP client.

Usage with llama-cli (CPU)

llama-cli --model /path/to/Godoter3.8-27B.Q4_K_M.gguf \
  --mmproj /path/to/Godoter3.8-27B-mmproj.gguf \
  -p "Explain how to use the RayCast2D node in Godot." -n 300

Training

• Base model: Qwen3.8-27B (loaded in 4-bit via the unsloth/Qwen3.8-27B 4-bit checkpoint)
• Fine-tuning method: QLoRA (r=32, alpha=64)
• Dataset: arunsrajan/godot-docs-dataset - 24,631 instruction/answer pairs, every row anchored to an official Godot Engine 4.x documentation page
• Sequence length: 4096
• Epochs: 3
• Framework: Unsloth 2026.10.5 on ROCm

Notes

• This fine-tune builds on the base model's general capabilities; general knowledge and reasoning remain strong, with Godot-related knowledge significantly improved.
• No benchmarks have been run yet; performance numbers will be added when available.

License

Apache 2.0 - same as the base Qwen3.8-27B weights.