bbidpa/Qwen2.5-Coder-0.5B-Flutter-direct

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 27, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

bbidpa/Qwen2.5-Coder-0.5B-Flutter-direct is a 0.5 billion parameter language model, fine-tuned from Qwen/Qwen2.5-Coder-0.5B, specifically designed for generating complete Flutter/Dart files from natural language goals. This model excels at single-shot code generation for Flutter applications, trained on 5 million tokens of goal-to-complete-file examples. It is part of a research study comparing direct code generation with iterative, diff-based approaches for small code models.

Loading preview...

Model Overview

bbidpa/Qwen2.5-Coder-0.5B-Flutter-direct is a specialized 0.5 billion parameter language model, fine-tuned from the Qwen2.5-Coder-0.5B base model. Its primary function is to generate a complete Flutter/Dart file in a single shot based on a natural-language goal.

Key Capabilities

  • Direct Flutter/Dart Code Generation: Generates entire Flutter/Dart files from a given natural language prompt, rather than iterative edits.
  • Optimized for Flutter Development: Specifically fine-tuned on 5 million tokens of flat goal-to-complete-file examples from the bbidpa/flutter-full-examples-v1 dataset.
  • Research Context: This model is a component of a comparative study investigating the effectiveness of direct code generation versus iterative, diff-based generation for small code models. Its companion model, Qwen2.5-Coder-0.5B-Flutter-steps, uses the same base model but is trained for iterative diff-based generation.

Usage and Integration

  • The model uses Qwen's native BPE tokenizer, extended with structural special tokens.
  • It follows a specific prompt format utilizing <GOAL>, <CODE>, <HISTORY>, and <OUTPUT> tags, where the model completes the content after <OUTPUT>.
  • Loading is straightforward using transformers' AutoModelForCausalLM and AutoTokenizer.

Related Work

This model is part of a broader research effort detailed in the paper "Diffs vs. Whole Files: An Empirical Comparison of Iterative Edit-Based and Direct Generation for Flutter/Dart Code Models" (arXiv:2609.05779).

Finetunes
1 models