revit-coder-14b

A domain-adapted 14B parameter language model fine-tuned for Autodesk Revit 2025+ API development. Generates production-grade C# code with correct transaction management, thread safety patterns, and crash-resistant architecture.

Built on Qwen3-14B using QLoRA. Available in F16 (full precision) and Q4_K_M (quantized) GGUF formats.

What This Model Does

This model is specialized for Revit API code generation and validation. It understands:

• Transaction hierarchy (Transaction, TransactionGroup, SubTransaction) and when to use each
• Thread safety with IExternalEventHandler and ExternalEvent patterns
• IFailuresPreprocessor for agentic/automated execution without modal dialog popups
• FilteredElementCollector efficiency patterns and filter chaining
• Batch APIs (NewFamilyInstances2, Delete(ICollection)) for high-performance operations
• Element lifecycle management and stale reference prevention
• Revit 2025+ specifics: .NET 8, C# 12, ElementId.Value (long), net8.0-windows targeting
• Worksharing safety checks and checkout patterns
• Electrical systems, PanelScheduleView handling, and unit conversion

Quick Start (Ollama)

Download the Q4_K_M GGUF (recommended for most hardware):

# Create a Modelfile
cat > Modelfile << 'EOF'
FROM ./revit-coder-14b-Q4_K_M.gguf

TEMPLATE """{{- if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
<think>

</think>

"""

SYSTEM "You are a Revit API specialist for Autodesk Revit 2025+ (.NET 8, C# 12). You generate production-grade C# code that is crash-resistant, transaction-safe, and optimized for agentic execution."

PARAMETER temperature 0.7
PARAMETER num_ctx 16384
PARAMETER stop "<|im_end|>"
PARAMETER repeat_penalty 1.1
EOF

# Create and run
ollama create revit-coder-14b -f Modelfile
ollama run revit-coder-14b "Write code to create 100 structural columns at specified coordinates"

Available Files

File	Size	Precision	VRAM Required	Use Case
revit-coder-14b-Q4_K_M.gguf	9 GB	4-bit quantized	~12 GB	Recommended for local use
revit-coder-14b-f16.gguf	26 GB	Full FP16	~28 GB	Maximum quality, requires 32GB+ RAM

Example Outputs

Transaction Safety

Prompt: "What's wrong with this code?"

for (int i = 0; i < walls.Count; i++)
{
    var t = new Transaction(doc);
    t.Start("Delete");
    doc.Delete(walls[i].Id);
    t.Commit();
}

Model response identifies:

• Anti-pattern: Transaction per loop iteration causes N regenerations instead of 1
• Missing using block (resource leak if exception thrown)
• Should use batch API: doc.Delete(wallIds) in a single Transaction
• Missing IFailuresPreprocessor (modal dialogs will block execution)

Batch API Knowledge

Prompt: "Create 100 structural columns at XYZ coordinates in Revit 2025"

The model generates code using doc.Create.NewFamilyInstances2() (the batch API) rather than looping with individual NewFamilyInstance() calls, resulting in 5-15x faster execution with a single regeneration.

Thread Safety

Prompt: "How do I call the Revit API from a background thread?"

The model correctly explains that Revit API is single-threaded and provides the IExternalEventHandler + ConcurrentQueue pattern, warning against Task.Run() or ThreadPool approaches that cause crashes.

Training Details

• Base model: Qwen3-14B
• Method: QLoRA (rank 64, alpha 128)
• Training data: Curated Revit API documentation, code patterns, and domain knowledge
• Domains: Revit C# API, IFC reasoning, APS/Forge schemas
• Epochs: 3
• Validation loss: 0.1586
• Hardware: RunPod GPU cluster

Recommended System Prompt

For best results, use a system prompt that encodes Revit transaction rules. A comprehensive system prompt is available in the companion repository under docs/.

Key rules to include:

• Transaction decision tree (dependent ops -> TransactionGroup, batch available -> batch API, otherwise -> single Transaction)
• Batch size limits (max 50 elements per Transaction, max 10 Transactions per TransactionGroup)
• Thread safety (IExternalEventHandler only, never background threads)
• Element lifecycle (store ElementId, never Element; re-fetch after commit)
• Always include IFailuresPreprocessor for agentic execution

Hardware Requirements

Format	Minimum RAM	Recommended	Inference Speed (M1 Max 64GB)
Q4_K_M	16 GB	32 GB	~11 tok/s
F16	32 GB	64 GB	~7 tok/s

Use Cases

• Revit plugin development: Generate boilerplate with correct transaction patterns
• Code review: Validate existing Revit API code against best practices
• Agentic AI: Serve as the code generation/validation layer in a Revit automation pipeline
• Learning: Understand Revit API patterns through conversational Q&A

Limitations

• Specialized for Revit 2025+ (.NET 8). May generate outdated patterns for Revit 2023/2024 (.NET Framework 4.8)
• Code generation, not execution. Generated code should be reviewed before running in production
• 14B parameter model. For complex multi-file architectural decisions, pair with a frontier model

License

Apache 2.0. Based on Qwen3-14B (Apache 2.0).

Author

Sanjay Chauhan -- GitHub | LinkedIn

Citation

@misc{chauhan2026revitcoder,
  title={revit-coder-14b: A Domain-Adapted LLM for Revit API Development},
  author={Chauhan, Sanjay},
  year={2026},
  url={https://huggingface.co/schauh11/revit-coder-14b}
}