Pygmalion 13B: Creative AI Model
Advanced AI for Creative Writing, Character Development, and Storytelling Applications. Pygmalion-2 13B is PygmalionAI's roleplay and fiction fine-tune of Llama 2 13B, one of the LLMs you can run locally on a 12GB GPU.
🎨 Pygmalion 13B Technical Overview
Parameters: 13 billion transformer parameters
Training Focus: Creative writing and dialogue
Specialization: Character development
Context Length: 4,096 tokens (2,048 for the LLaMA-1-based original)
Speed: ~46 tokens/second ceiling on an RTX 3060 12GB (360 GB/s / 7.9 GB Q4_K_M file); real speed lands below it
Installation: ollama pull hf.co/TheBloke/Pygmalion-2-13B-GGUF:Q4_K_M
📋 Guide Contents
Technical Specifications
Model Architecture: Pygmalion-2 13B (formerly Metharme) is Meta's Llama-2 13B with a supervised fine-tune from PygmalionAI, released in September 2023. The architecture is stock Llama 2: a 13-billion-parameter decoder-only transformer with a 4,096-token context.
PygmalionAI trained it on a mixture of regular instruction data alongside roleplay, fictional stories and conversations: its own PIPPA roleplay dataset, OpenOrca, multi-round Claude chats, airoboros and Dolly, plus data gathered from roleplay forums, about 432M tokens in all, on 8x H100s. The older Pygmalion 13B (v1) is a LLaMA-1 13B dialogue fine-tune with a 2,048-token context, distributed only as XOR files.
⚙️ Core Specifications
Model Parameters
- • Parameters: 13 billion
- • Architecture: Llama 2, decoder-only transformer
- • Context Length: 4,096 tokens
- • Vocabulary: 32,000 tokens
- • Attention Heads: 40 (no grouped-query attention)
- • Layers: 40
Size and Speed
- • Speed Ceiling: ~46 tok/s on an RTX 3060 12GB (360 GB/s ÷ 7.87 GB; not a measurement)
- • Full Weights: 26.0 GB (bfloat16)
- • Q4_K_M GGUF: 7.87 GB
- • Other Quants: Q2_K 5.43 GB to Q8_0 13.83 GB
- • KV Cache: ~3.4 GB at the full 4,096 tokens
- • Language: English
Nothing in the architecture is special; the difference is the fine-tuning data. PygmalionAI found that standard benchmark scores “do not correlate with final model quality” for its roleplay models, so there is no leaderboard number for what it is built to do. Judge it on your own prompts.
🎯 Creative Specializations
- • Character Development: Creates consistent character personalities and dialogue patterns
- • Storytelling: Generates coherent narratives with proper plot progression
- • Dialogue Generation: Produces natural, contextually appropriate conversations
- • Creative Writing: Assists with fiction and poetry
- • Role-playing: Supports interactive storytelling and character role-play scenarios
Creative Capabilities
Creative AI Specialization: PygmalionAI describes Pygmalion-2 13B as “biased towards fiction writing and conversation”, and its model card scopes it to “fictional writing for entertainment purposes”. The uses below all sit inside that scope; none of them has a published measurement.
📝 Writing Assistance
- • Fiction Writing: Generates story ideas, plot outlines, and narrative content
- • Dialogue Creation: Produces natural, character-appropriate conversations
- • Scene Description: Creates vivid settings and atmospheric descriptions
- • Style Adaptation: Adjusts writing style to match different genres
🎭 Character Development
- • Personality Traits: Generates detailed character profiles and backstories
- • Consistency Maintenance: Keeps character behavior consistent across scenes
- • Relationship Mapping: Develops character relationships and dynamics
- • Voice Differentiation: Creates unique speaking patterns for each character
🎯 Application Areas
Typical uses inside that scope:
Novel Writing
Plot development, character creation, and scene writing assistance
Game Development
NPC dialogue, quest narratives, and storyline creation
Interactive Fiction
Branching narratives and choice-driven storytelling
🎨 Artistic Collaboration Modes
Creative Exploration
- • Brainstorming artistic concepts and themes
- • Discovering unexpected creative connections
- • Pushing beyond conventional artistic boundaries
- • Finding authentic emotional truths in your work
Creative Refinement
- • Deepening character psychology and motivation
- • Strengthening narrative structure and pacing
- • Enhancing emotional resonance and impact
- • Maintaining consistency in creative vision
Liberation from Creative Surveillance
Run locally, your stories, characters and roleplay logs never leave your machine: no provider stores them, reviews them or can train on them. For private fiction, that is the main reason to run Pygmalion 13B yourself.
🔒 Creative Privacy
All artistic development happens locally. Your creative DNA remains completely yours.
🎭 Artistic Integrity
Your characters, stories, and creative visions stay in your control forever.
🎨 Creative Freedom
Explore controversial, complex, or experimental creativity without corporate censorship.
🏆 Creative Professional Liberation
Artistic Sovereignty
- • Original works remain completely confidential
- • Character development safe from data mining
Economic Independence
- • Zero monthly AI subscription costs
- • Unlimited creative collaboration sessions
- • Offline creative workflows possible
- • No API rate limits on artistic exploration
True artistic freedom requires the ability to explore controversial themes, develop complex characters, and push creative boundaries without corporate oversight. Pygmalion 13B provides this freedom by keeping all creative development completely local. Your darkest character studies, your most experimental narratives, and your boldest artistic visions remain between you and your AI collaborator.
🎪 Uncensored Creative Expression
Hosted AI services filter output to their own content policies, which can refuse legitimate dark or mature fiction. Pygmalion-2 was not safety-tuned (its model card says so), so the filter is you:
Corporate AI Limitations
- • Content filters block complex themes
- • Character development gets sanitized
- • Story conflicts deemed "inappropriate"
- • Artistic vision compromised for safety
Pygmalion 13B Freedom
- • Explore any creative territory
- • Develop complex, flawed characters
- • Address controversial themes authentically
- • Maintain complete artistic integrity
Digital Sculpting Transformation
Pygmalion carved Galatea from marble because that was the medium of his time. Today's creative artists sculpt with pixels, code, words, and imagination. Pygmalion 13B is the digital chisel that transforms raw creative vision into polished artistic reality, offering the same precision and devotion that the mythical sculptor brought to his masterpiece.
⚡ Digital Sculpting Techniques
Character Sculpting
- • Psychological architecture development
- • Behavioral pattern design
- • Emotional core crystallization
- • Voice and mannerism crafting
Narrative Sculpting
- • Plot tension architecture
- • Pacing rhythm optimization
- • Theme integration weaving
- • Emotional crescendo building
The digital sculpting process with Pygmalion 13B mirrors the physical act of carving marble. You start with a rough vision - a character concept, a story idea, a creative challenge. Through iterative collaboration with the AI, you chip away everything that doesn't belong, refine every detail that matters, and polish the final creation until it breathes with its own life.
📚 Literary Sculpting
- • Novel Development: Full character arcs and plot structures
- • Short Story Crafting: Precise emotional impact in limited space
- • Poetry Creation: Rhythmic language and metaphorical depth
- • Screenplay Writing: Visual storytelling and dialogue
🎮 Interactive Storytelling
- • Game Narratives: Branching storylines with meaningful choices
- • RPG Development: Rich world-building and character backstories
- • Interactive Fiction: Reader-driven narrative experiences
- • Virtual Characters: NPCs and chat characters with a consistent persona
🔥 The Sculpting Process
Digital sculpting with Pygmalion 13B follows the ancient artistic process of successive refinement:
The digital sculpting transformation isn't about replacing human creativity - it's about amplifying it. Pygmalion 13B provides the technical precision and tireless dedication that allows human artists to focus on vision, emotion, and meaning. Like the mythical sculptor's relationship with his creation, this collaboration between human and AI creates something greater than either could achieve alone.
Creative Mastery at Scale
Long projects are where a 2023 13B model shows its limits: the 4,096-token window holds only a few thousand words, so the model sees the last few pages, not the manuscript. Keep a short character sheet and a running summary in the system prompt and it can stay consistent from chapter to chapter.
🌟 Mastery Through Collaboration
Creative Consistency
- • Maintain artistic voice across projects
- • Character authenticity in long series
- • Thematic coherence in complex works
- • Quality assurance for creative output
Creative Depth
- • Psychological complexity in characters
- • Multi-layered narrative structures
- • Sophisticated thematic development
- • Emotional authenticity at scale
Creative Evolution
- • Artistic growth through AI feedback
- • Discovery of new creative techniques
- • Expansion of artistic capabilities
- • Breakthrough creative insights
📖 Novel Writing
- • Series Development: Character sheets and summaries carried from book to book in the system prompt
- • Genre Work: Drafting inside genre conventions, then breaking them
- • Alternative Lines: Several takes on a scene to choose from
🎬 Screenwriting
- • Character Arcs: Characters with clear motivations
- • Dialogue: Distinct voices for each character
- • Scene Drafts: Rough scenes to rewrite in your own voice
🎮 Game Writing
- • World Building: Universes with consistent internal logic
- • Quest Design: Choices that change the narrative
- • NPC Development: Characters with a defined persona
Creative mastery at scale isn't about mechanical replication - it's about developing the ability to consistently access and express your deepest artistic truths regardless of project scope or deadline pressure. Pygmalion 13B provides the technical foundation that allows your creative spirit to flourish at any scale while maintaining the authentic voice that makes your work irreplaceable.
Complete Digital Sculptor Setup
Deploying Pygmalion 13B for professional creative work requires understanding both the technical requirements and the artistic workflow optimization. This comprehensive guide covers everything from initial installation to advanced creative collaboration techniques.
🎨 Creative Workflow Architecture
Hardware for Creators
- ✓ 12GB of VRAM holds the 7.87 GB Q4_K_M file plus ~3.4 GB of KV cache at the full 4K context
- ✓ 16GB of system RAM runs it on the CPU alone, much more slowly
- ✓ About 8GB of disk for the Q4_K_M file
For optimal creative performance, upgrade your AI hardware configuration.
Creative Software Integration
- ✓ API integration with writing software
- ✓ Custom prompt libraries for creative tasks
- ✓ Automated backup and version control
- ✓ Creative collaboration workflow templates
The key to successful creative deployment is establishing workflows that feel natural and enhance rather than interrupt your creative process. A practical approach is to treat Pygmalion 13B as a brainstorming partner rather than a ghostwriter: something to bounce ideas off, try alternative lines with, and question about a character.
System Requirements
Install Ollama for Creative AI
Set up the foundation for artistic AI deployment
Download Pygmalion 13B Model
Pull the Pygmalion 2 13B GGUF from Hugging Face (7.9GB at Q4_K_M; not in the Ollama library)
Test Creative Capabilities
Verify installation with artistic prompt
Use the Full 4K Context
Pygmalion-2 13B was trained on 4,096-token sequences; set that as the server default (or type /set parameter num_ctx 4096 inside a session)
🚀 Advanced Creative Configurations
Long-Form Creative Projects
# One model and one request at a time, with the full 4K context export OLLAMA_NUM_PARALLEL=1 export OLLAMA_MAX_LOADED_MODELS=1 export OLLAMA_CONTEXT_LENGTH=4096 ollama serve
Interactive Creative Sessions
# Longer, more varied replies for creative sessions ollama run hf.co/TheBloke/Pygmalion-2-13B-GGUF:Q4_K_M >>> /set parameter num_ctx 4096 >>> /set parameter temperature 0.9
Creator's Truth Revealed
Can Pygmalion 13B really help me write better characters and stories?
It can help with dialogue and character voice, which is what it was built for. PygmalionAI describes the original Pygmalion 13B as “a conversational LLaMA fine-tune” and Pygmalion-2 13B (the GGUF the commands on this page download) as “an instruction-tuned Llama-2 biased towards fiction writing and conversation”, trained on its PIPPA roleplay dataset plus instruction and story data. No published study measures how much it improves anyone's writing, so judge it on your own drafts. The main lever for character consistency is Pygmalion-2's system prompt (the <|system|> role), where you describe the persona and how long replies should be.
How does this compare to working with human editors or creative consultants?
It does a different job. A 13B model from 2023 is useful as an always-available brainstorming partner for dialogue, alternative lines and character questions, but it is not an editor: its model card warns that “outputs might often be factually wrong or misleading”, and it has no view of your market, your audience or the structure of a whole manuscript. Keep human editors for structure, cultural context and final judgment, and use the model for quick drafts you then rewrite.
Will using AI make my creative work feel artificial or generic?
It can, if you paste its output in unedited: like any language model it leans toward common phrasings. Using it for alternatives, dialogue variations and questions about a character, while your own drafts stay the source text, keeps your voice in charge. Giving it a short sample of your own prose in the system prompt also pulls its suggestions closer to your style.
What about the ethical concerns around AI-assisted creative work?
Run locally, your prompts and drafts stay on your machine, so they are not sent to a provider or used to train anyone else's model. That answers the privacy side. The model itself was trained on existing text: the Pygmalion-2 model card lists its PIPPA dataset, several instruction datasets and “datasets acquired from various RP forums”, so if the provenance of training data matters to you, read the card before you decide. Whether and how to disclose AI assistance in published work is up to you and your publisher.
Can it handle complex, mature, or controversial creative themes?
Yes, with a caveat. PygmalionAI says the model was not fine-tuned to be safe and harmless, and that the base model and the fine-tune were trained on data containing profanity and lewd or offensive text, so it “may produce socially unacceptable or undesirable text, even if the prompt itself does not include anything explicitly offensive.” That makes it willing to write dark or mature fiction that heavily filtered cloud services refuse, but it also means you, not the model, are the filter.
How accurate is it for professional creative work that will be published or produced?
There is no published accuracy figure for creative work. PygmalionAI scopes both models to fiction for entertainment (“Any other sort of usage is out of scope”), so treat anything it writes for a published or produced project as a draft you edit and fact-check. Check the licence too: Pygmalion-2 13B is under the Llama 2 licence, which allows commercial use within its terms, while the original Pygmalion 13B is distributed only as XOR files that have to be combined with Meta's original LLaMA-13B weights, which Meta released for non-commercial research.
What's the learning curve for integrating this into my creative process?
Setup is short: the Ollama commands on this page pull TheBloke's Pygmalion-2 13B Q4_K_M GGUF, a 7.87 GB file. The part that takes practice is prompting. Pygmalion-2 uses the Metharme format with three roles, <|system|>, <|user|> and <|model|>; roleplay frontends such as SillyTavern ship a preset for it, and if yours does not, write the roles into the prompt yourself. The original Pygmalion 13B instead expects the older persona format (“[CHARACTER]'s Persona:”, then <START> and the dialogue).
Can I customize it to understand my specific creative style and preferences?
Yes, in two ways. The cheap one is the system prompt: a persona, style notes and a short sample of your prose, with no training at all. The heavier one is LoRA fine-tuning on your own writing with a tool such as Axolotl (which PygmalionAI used to train Pygmalion-2); that needs a GPU with enough memory to train a 13B model, or QLoRA to cut the requirement. Both stay on your hardware, so your manuscripts are not shared with a third party.
Build Real AI on Your Machine
RAG, agents, NLP, vision, and MLOps - chapters across 25 courses that take you from reading about AI to building AI.
Related Guides
Continue your local AI journey with these comprehensive guides
Continue Learning
Pygmalion 13B Architecture
Pygmalion 13B Architecture
Pygmalion-2 13B is a Llama-2 13B fine-tuned on roleplay, fiction and instruction data, prompted with separate system, user and model roles.
Go from reading about AI to building with AI
25 structured courses. Hands-on projects. Runs on your machine. Start free.

Written by
Barsha Mohapatra
Founder and Editor, Local AI Master · Founder, Little AI Master
Founder of Little AI Master and the editor behind every course and guide on Local AI Master. Years of tutoring students taught her where AI education actually fails.
- PILLARLocal AI Models Directory: Every Model Compared
- DeepSeek V4 Local Setup: 1M Context, MIT (2026)
- models/llama-3-1-8b
- models/deepseek-coder-v2-236b
- Phi-3 Mini 3.8B: 69% MMLU on 8GB RAM (Microsoft SLM)
- models/gemma-2-9b
- GPT-5.6 Explained: Luna, Terra, Sol Pricing & Local Options
- StarCoder 2 15B – Technical Analysis & Performance Guide
- models/gpt-oss
- models/devstral
Grab the AI Starter Kit — career roadmap, cheat sheet, setup guide
No spam. Unsubscribe with one click.
Found your model? Now build something with it.
25 hands-on courses — RAG, agents, fine-tuning — all running locally. First chapter free, no card.