arbazsiddiqui/Ozan-v1-12B

TEXT GENERATIONConcurrency Cost:1Model Size:12BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 25, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

Ozan-v1-12B is a 12 billion parameter creative-writing model developed by arbazsiddiqui, fine-tuned from Mistral-Nemo-Instruct-2407 with a 32768 token context length. It is specifically optimized to produce long-form fiction that avoids common "LLM tells" or "slop," achieving the lowest slop metric among runnable 12B models tested. This model excels at generating high-quality, natural-sounding prose for creative writing applications.

Loading preview...

Ozan-v1-12B: A Low-Slop Creative Writing Model

Ozan-v1-12B, developed by arbazsiddiqui, is a 12 billion parameter model fine-tuned from Mistral-Nemo-Instruct-2407 with a 32K context length. Its core innovation is the deliberate reduction of "slop"—overused, generic phrases common in AI-generated text—making its output sound more human and less like an LLM. This is achieved through a unique training methodology.

Key Capabilities & Training

  • Low Slop Generation: Ozan-v1-12B is specifically trained and measured to minimize common LLM tells, achieving the lowest slop metric among comparable 12B models on the EQ-Bench Creative Writing v3 harness.
  • Creative Writing Focus: It is designed for long-form fiction, producing prose that is fluent, coherent, and strong in voice and tone.
  • Hybrid Training: The model underwent Supervised Fine-Tuning (SFT) using QLoRA on ~17k curated, low-slop examples from non-GPT models and human writing prompts. This was followed by Gutenberg anti-slop DPO (Direct Preference Optimization), which preference-tuned the model towards real public-domain book prose.
  • Performance: While excelling in slop reduction (5.30 slop, 3.21 repetition), it also shows competitive craft scores (58.8/100) against stronger low-slop 12B finetunes, significantly improving over its base model.

Use Cases

  • Long-form Fiction: Ideal for generating stories, chapters, or other extensive creative narratives.
  • Character Dialogue: Capable of producing distinct character voices, as demonstrated in provided samples.
  • Avoiding AI-isms: Suitable for applications where the goal is to produce text that is indistinguishable from human writing, particularly in creative contexts.

Technical Details

  • Base Model: Mistral-Nemo-Instruct-2407.
  • Quantization: Available in GGUF quants (Q4_K_M, Q5_K_M, Q6_K, Q8_0) with importance matrix (imatrix) for optimal quality.
  • Recommended Sampler Settings: Uses temperature 0.7, min_p 0.1, and DRY (Dynamic Repetition control) for clean, long outputs.

Finetunes
2 models