import { pipeline } from '@xenova/transformers';

// Create pipeline
const generator = await pipeline('text-generation', 'SpongeBOB9684/qwen3.5-0.8b-mermaid-generator-onnx', {
    dtype: 'q4',  // quantized model
    device: 'webgpu',
});

// Generate
const prompt = 'Create a flowchart for a simple login process';
const messages = [
  { role: 'system', content: 'You are a Mermaid diagram code generator. Output ONLY valid Mermaid code.' },
  { role: 'user', content: prompt },
];

const output = await generator(messages);
console.log(output);

Performance

Model Size: ~1.3 GB (FP16 quantized)
Load Time: < 5 seconds on typical browsers
Inference Speed: ~15-30 tokens/second (depends on hardware)
Memory: ~1.3 GB GPU memory (with quantization)

Advantages

FP16 Quantization:
- ~50% smaller model size
- Faster inference with minimal quality loss
- Lower memory usage for browser deployment
ONNX Format:
- Optimized for web inference
- Cross-platform compatibility
- Direct loading in Transformers.js

Limitations

FP16 quantization may cause slight precision differences compared to FP32
Best results with clear, specific prompts
Limited to Mermaid syntax (not general diagram description)

License

Apache 2.0

Acknowledgments

Base model: Qwen3.5-0.8B by Alibaba Cloud
ONNX conversion: Optimum library
Dataset: SpongeBOB9684/mermaid-text-to-diagram
Training framework: Unsloth