Overview
JanusCoder-14B is a 14 billion parameter text model from InternLM, based on Qwen3-14B. It is part of the JanusCoder suite, which aims to create a unified visual-programmatic interface for code intelligence. The model is trained on JANUSCODE-800K, the largest multimodal code corpus to date, which includes data for generating standard charts, complex interactive Web UIs, and code-driven animations.
Key Capabilities
- Unified Visual-Programmatic Interface: Handles diverse visual-programmatic tasks, generating code from text, visual inputs, or a combination.
- Flexible Content Generation: Capable of creating data visualizations and interactive front-ends.
- Program-Driven Editing: Supports precise editing of visual effects and complex animation construction.
- Multimodal Code Corpus Training: Leverages JANUSCODE-800K for comprehensive code intelligence.
Performance
JanusCoder-14B demonstrates strong performance across various code intelligence benchmarks, including:
- PandasPlotBench: Achieves 86, outperforming Qwen3-14B (78) and LLaMA3-8B-Instruct (69).
- ArtifactsBench: Scores 41.1, surpassing Qwen3-14B (36.5) and GPT-4o (37.9).
- DTVBench (Manim): Scores 8.41, higher than Qwen3-14B (6.63) and LLaMA3-8B-Instruct (4.92).
Good For
- Developers needing to generate code for data visualizations.
- Creating interactive web UIs from high-level descriptions.
- Tasks involving program-driven manipulation of visual elements and animations.