Kezmark/Mordant-12B-Think

TEXT GENERATIONConcurrency Cost:1Model Size:12BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 18, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

Kezmark/Mordant-12B-Think is a 12 billion parameter, Mistral-Nemo-2407-based language model fine-tuned for AI image generation composition. It excels at producing detailed, narratively rich compositions with chain-of-thought reasoning, frequently self-critiquing decisions and explaining compositional choices. This model is optimized for generating spatially structured image prompts and complex multi-element scenes requiring narrative-level spatial reasoning.

Loading preview...

Mordant-12B-Think: Flagship Compositional AI

Kezmark/Mordant-12B-Think is the flagship model of the Mordant family, a 12 billion parameter model based on Mistral-Nemo-2407. It is specifically fine-tuned for AI image generation composition, focusing on creating detailed and narratively rich prompts.

Key Capabilities

  • Advanced Chain-of-Thought Reasoning: The model demonstrates genuine analytical depth, self-critiquing decisions and explaining why specific elements are placed as they are, often with a "sharp, caustic, professional mentor" personality during the reasoning phase.
  • Detailed Spatial Structuring: Produces compositions that go beyond simple descriptions, constructing deliberate meaning through compositional choices and spatial relationships.
  • Extensive Training: Trained on a curated dataset of 7,203 highly detailed compositions across various tasks, genres, art styles, and emotional contexts.
  • Art Style Versatility: Explicitly trained on over 40 distinct art styles, from "16-bit Pixel Art" to "Watercolor" and "Vintage Pulp Fiction", ensuring diverse output capabilities.

Good For

  • Generating detailed, spatially structured image prompts from short descriptions.
  • Hero-tier composition work where quality and narrative depth are paramount.
  • Creating complex multi-element scenes requiring advanced spatial reasoning.
  • Enhancing existing prompts with deep compositional treatment.

Comfy-UI Integration

A custom MordantPromptEnhancer Comfy-UI node is included, which handles GGUF loading, GPU VRAM auto-tuning, tokenization, and chain-of-thought parsing, simplifying local inference.