aifeifei798/QiMing-LongWriter-20B-MXFP4
The QiMing-LongWriter-20B-MXFP4 is a 20-billion-parameter Mixture-of-Experts (MoE) language model developed by aifeifei798, specifically engineered as a "Narrative Engine" for creative writing. It features MXFP4 quantization for efficient inference on consumer hardware and boasts a theoretical context length of 131,072 tokens, though it demonstrates professional-grade creative capabilities within an 8,192-token window. This model excels at long-form narrative generation, screenplay outlining, and creative concept development, utilizing a unique "Process-Oriented Overfitting" training methodology to internalize creative workflows rather than just memorizing text.
Loading preview...
QiMing-LongWriter-20B-MXFP4: A Narrative Engine for Creative Writing
QiMing-LongWriter-20B-MXFP4, developed by aifeifei798, is a 20-billion-parameter Mixture-of-Experts (MoE) language model designed as a "Narrative Engine" for professional creative writing. It leverages MXFP4 quantization, allowing it to run efficiently on consumer-grade hardware with 16GB of VRAM. While possessing a theoretical context length of 131,072 tokens via YaRN Rope Scaling, its true strength lies in its demonstrated professional-grade creative capabilities within a practical 8,192-token window.
This model's unique "Process-Oriented Overfitting" training paradigm teaches it the 'how' of creation, focusing on a "Writer's Long-form Logic" framework rather than just learning from massive text corpora. This methodology enables it to maintain long-range logical coherence and structure in creative tasks, making it distinct from models that rely solely on brute-force memory.
Key Capabilities
- Long-form Narrative Structure: Excels at generating outlines, chapter summaries, and first drafts for various genres.
- Screenplay & Outline Construction: Capable of creating detailed frameworks for multi-episode series and feature films.
- Game World-Building: Develops rich lore, backstories, and narrative questlines.
- Brainstorming & Idea Expansion: Transforms vague ideas into structured and imaginative narrative paths.
- Efficient Inference: MXFP4 quantization allows operation within 16GB of memory.
- Configurable Reasoning: Offers low, medium, and high reasoning levels for varied task requirements.
- Agentic Capabilities: Supports function calling, web browsing, Python code execution, and Structured Outputs.
Good For
- Creative Writers & Authors: Ideal for generating novel outlines, story drafts, and character arcs.
- Screenwriters & Game Designers: Useful for developing detailed plot structures, world lore, and quest designs.
- Developers requiring efficient creative generation: Its MXFP4 quantization and MoE architecture make it suitable for deployment on more accessible hardware.
- Experimentation and Customization: Permissive Apache 2.0 license and fine-tunability support broad use cases.
Limitations
- Hardware Dependency: While efficient, utilizing the full 131k context requires significant VRAM; practical context lengths (e.g., 8k, 16k, 32k) are recommended for most hardware.
- Non-Factual Nature: Optimized for creativity, not factual accuracy; outputs may be inaccurate or fabricated.
- Prompt Quality Dependent: Requires clear, structured prompts to unlock its full potential.