Janstnkel/Godoter-27B

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 7, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Janstnkel/Godoter-27B is a 27-billion-parameter Qwen3.6-27B fine-tune, specifically optimized for generating Godot 4 and GDScript code. This model excels at producing idiomatic Godot 4 code, handling complex multi-file systems, and consistently avoiding Godot 3 API confusion or framework drift. It is designed for developers requiring reliable and accurate Godot 4 GDScript generation.

Loading preview...

Godoter-27B: Specialized Godot 4 & GDScript Model

Godoter-27B is a 27-billion-parameter model developed by Janstnkel, fine-tuned from the Qwen3.6-27B base model using QLoRA. Its primary purpose is to provide highly reliable and accurate code generation for Godot 4 and GDScript, specifically designed to prevent common issues like API version confusion or drifting to other programming frameworks.

Key Capabilities

  • Idiomatic Godot 4 GDScript: Generates correct code for nodes, signals, resources, @export/@onready, await, and typed code.
  • Complex System Building: Capable of creating complete, multi-file systems such as inventory, FSM AI, save/load, dialogue, multiplayer, shaders, and combat mechanics.
  • Version Consistency: Reliably uses Godot 4 APIs, avoiding deprecated Godot 3 functions like move_and_slide() or connect().
  • Framework Anchoring: Stays focused on Godot, preventing responses in other languages or frameworks like HTML/CSS/JS, Python, or gRPC.

Performance & Training

Benchmarking against its base Qwen3.6-27B model, Godoter-27B achieved 100% on easy Godot 4 migration tasks and 96% on advanced Godot 4 tasks, significantly outperforming the base model's 80% and 68% respectively. This improvement is largely due to its specialized training on approximately 18,000 examples, including Godot 4 API references, Q&A over tutorials, and 666 complete multi-file Godot 4 systems. The model was trained using QLoRA with r=16, alpha=16, and adamw_8bit on a 4-bit base, with the adapter merged to 16-bit.

Limitations

As a derivative of Qwen3.6-27B, its general reasoning capabilities are largely inherited from the base. The fine-tune primarily enhances Godot 4 reliability. Its knowledge is bounded by its training data, meaning very recent Godot 4.x changes might not be reflected. Users should always review generated code for architectural quality and correctness before deployment.