youngkim/qwen35-4b-text2svg-ar-mtp

VISIONConcurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 26, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The youngkim/qwen35-4b-text2svg-ar-mtp is a 4.5 billion parameter Qwen3.5-based model developed by youngkim, specifically designed for Text-to-SVG generation. This model incorporates multi-token prediction (MTP) for speculative decoding, significantly accelerating output generation. It is optimized for structured slide briefs to SVG, offering up to 2.44x speedup in output token generation compared to baseline autoregressive decoding.

Loading preview...

Overview

This model, youngkim/qwen35-4b-text2svg-ar-mtp, is a specialized 4.5 billion parameter Qwen3.5-based checkpoint for Text-to-SVG generation. It integrates Multi-Token Prediction (MTP) for speculative decoding, building upon the youngkim/qwen35-4b-text2svg-ar base model. The MTP module was post-trained on 5,000 on-policy examples over 3 epochs, with only 12 MTP-specific tensors updated while the autoregressive body remained frozen.

Key Capabilities & Performance

  • Accelerated Decoding: Benchmarks on an NVIDIA H100 GPU show significant speed improvements. For instance, MTP-4 mode achieves a 2.44x median speedup in output token generation (561.7 tok/s) compared to the baseline (231.9 tok/s), with MTP-3 offering a 2.27x speedup.
  • High Validity: Maintains high SVG output validity, with 97 out of 100 benchmarked prompts producing valid SVG in MTP-4 mode.
  • Flexible Deployment: Supports both MTP-enabled speculative decoding via vLLM (recommended for production) and standard autoregressive decoding using the Hugging Face Transformers library.

Use Cases & Considerations

  • Specialized Text-to-SVG: This model is specifically engineered for converting structured text inputs, such as slide briefs, into SVG graphics. It is not a general-purpose instruction model.
  • Production Readiness: For production, it's recommended to implement structural validation and a fallback mechanism (e.g., retrying with MTP-3 if MTP-4 output is truncated or invalid) due to the nature of speculative decoding.
  • Hardware Optimized: Benchmarks were conducted on NVIDIA H100, indicating its suitability for high-throughput environments.

Finetunes
1 models