GrimoireScribe will ship with a purpose-built 6.87 GB writing model that runs entirely on your hardware. Built on Google's Gemma 4 12B, then fine-tuned on 2,000 hand-crafted creative-writing examples drawn from every GrimoireScribe preset. A 256,000-token context window. This is not a generic chatbot stuffed into a writing app. This is a model built for novelists.
Our model is not trained from scratch. It stands on Google's Gemma 4 12B, then adds a layer that is entirely ours.
Google's Gemma 4 is a 12-billion-parameter open model built on the same research that powers Gemini. Its hybrid attention design interleaves local and global layers, which keeps it fast and memory-light for its size: it runs quicker than a typical 12B and carries a 256,000-token context window. Just as important, its pretraining was filtered for safety, so responsible content boundaries are built into the foundation rather than bolted on. This gives our model its raw intelligence, reasoning, and language understanding.
Our layer. We fine-tune Gemma 4 on 2,000 hand-crafted creative-writing examples: 1,500 built through every one of our 15 writing presets, plus 500 freeform prose samples. No public-domain filler, no code, no academic text. Pure creative fiction. This is where the model learns our specific craft philosophy: show don't tell, subtext-driven dialogue, preset-native style. We fine-tune the Q4_K_M quantization of Gemma 4 12B directly: the same 6.87 GB file you download. What we train is exactly what you run.
The result: a compact model that carries Gemma 4's broad intelligence and GrimoireScribe's specific writing craft baked into its weights. It will download once through the app, then live on your machine permanently. No internet required. No subscription. No cloud calls.
Context window: 256,000 tokens, roughly 192,000 words. For nearly any novel, that means the model can read your entire manuscript in a single pass.
Our fine-tuning layer is 2,000 hand-crafted training examples, and every single one is modern creative fiction. No public-domain padding. No nineteenth-century novels scraped to inflate a number. No code, no academic papers, no generic web text. We made that choice deliberately. A model fed centuries-old prose learns to write like it: archaic cadence, dated vocabulary, sentence rhythms that read as costume rather than craft. We wanted a model that writes the way novelists write now.
Of those 2,000 examples, 1,500 were built through every one of our 15 writing presets. The remaining 500 are freeform prose samples that keep the model's range wide. What makes them work is not their quantity. It is what is inside them.
Golden Examples are hand-crafted training samples that encode the exact writing quality we want the model to produce. Each one is a pair: an instruction ("Write a scene where...") and a prose output that meets GrimoireScribe's standards. All 1,500 examples were built against our writing style presets, ensuring the model learns our craft philosophy from the inside out.
Think of it this way: Gemma 4 gives the model its intelligence and its safety foundation. The Golden Examples teach it how GrimoireScribe prose works specifically.
Physical anchoring, sensory triggers, backstory revealed through action rather than exposition. The model learns to convey emotion through what characters do, not what they feel.
Natural conversation where characters talk around what they mean. Minimal tags. What goes unsaid carries as much weight as what is spoken.
Wrongness-first worldbuilding. The strange is presented as fact, not explained. Magic systems conveyed through constraint and consequence, not exposition dumps.
Internal states revealed through behavior, micro-decisions, and physical response. No labeling emotions. The reader infers the psychology from what the character does.
Register shifts between intimacy and grandeur. Sentence structure that mirrors emotional state. The model learns to match tone to moment, not default to one register.
Violence with consequence. Intimacy with honesty. Moral complexity without judgment. The model does not shy away from the parts of fiction that make stories real.
Rituals shown as automatic habit, not anthropological lecture. Spiritual practices embedded in character behavior. Worlds feel lived-in because the details are woven, not announced.
Environmental logic and unremarked details that pay off later. The model learns to plant seeds without highlighting them. Subtlety is the skill being trained here.
Not just generation. The model also learns to take weak prose and improve it: tightening language, removing crutch words, and elevating passages to match your chosen preset.
Most local AI models ignore detailed style instructions. You write a careful system prompt describing your voice, your tone, your rules. The model reads it, nods politely, and produces the same generic output it always does. This happens because your prompt conflicts with what the model learned during training. The trained behavior wins.
GrimoireScribe's model was trained on the presets. The writing style rules are baked into the model's weights, not fighting against them. When you select a preset and write a prompt, the system prompt and the model's training reinforce each other instead of competing. The prompt can focus on the creative direction of your scene because the craft principles are already handled.
The result: a local model that actually follows your style instructions. Not perfectly. Not at the level of a frontier cloud model like Claude or GPT. But meaningfully, noticeably better than any off-the-shelf model trying to parse a style guide from a system prompt alone.
A writing model is only as good as the context it receives. Most AI writing tools either dump your entire manuscript into the prompt (expensive, slow, often truncated) or give the AI nothing and hope for the best. GrimoireScribe takes a different approach.
The app builds a structured knowledge system for every book: a fast structural index of your chapters, scenes, and characters. A full-text search layer that can find any passage in your manuscript instantly. AI-generated summaries of character arcs and chapter events that update automatically as you write.
When you ask the AI anything, a query router determines what context is needed and assembles it. A question about a character pulls their arc summary and scene appearances. A question about your magic system searches your prose and world notes. A request to write a new scene loads the prior scene, your style guide, character profiles, and relevant world-building notes.
This architecture is the same whether you use cloud providers or Offline Mode. The knowledge layers, full-text search, and query routing are all local. They run on your machine, in milliseconds, at zero cost. The only thing that changes between online and offline is which model generates the response.
Chapter titles, scene lists, word counts, character appearances. Built from your files in milliseconds. No AI needed. Always current.
Find any passage in your manuscript by content, not just title. "Find scenes about the magic system" returns matching scenes instantly. Runs locally via SQLite.
Your lore bibles and research notes are searchable by keyword locally. Semantic search is available through optional cloud or local embedding models for writers with large world-building collections.
Fiction needs room to go to dark places. Dark fantasy needs violence with weight. Romance needs intimacy with honesty. Grimdark needs moral ambiguity without the AI stopping to teach a lesson. Horror needs to unsettle without pulling its punches. Our model was trained for all of it.
It also has boundaries, and we would rather be straight with you about them than pretend otherwise. During evaluation we tested fully uncensored, "abliterated" models. Abliteration does not simply relax creative restrictions. It strips every guardrail a model has, including the ones that exist to stop the worst imaginable content. We will not ship a model like that, and we will not build a writing studio on top of one. The Gemma 4 foundation keeps its safety training intact, and fine-tuning it on creative prose does not undo that.
Here is what that means in practice.
This is the difference we want to stand behind: our model writes well because it was trained on good writing, not because its safety was removed. That is the opposite of the abliterated-model corner of the internet, and it is a line we are glad to hold.
This is not a ship-and-forget model. As the product evolves and we listen to community feedback, there may be future improvements to the local model, released for free through software updates. Nothing about your writing is ever sent to us. Improvements come from our own continued training, never from your data. Your local AI can get better over time at no extra cost.
Improvements will be documented on our blog as we release them: what changed, why it changed, and what to expect. We believe in showing our work.
Our roadmap includes expanding genre-specific training (more literary fiction, more thriller and mystery craft, more nonfiction techniques), improving edit mode accuracy, and refining the model's ability to match established voice across longer passages. The 2,000 Golden Examples we ship with (1,500 preset-guided plus 500 free-prose) are the beginning, not the ceiling.
We believe in honesty, so we measured the gap instead of hand-waving it, and we will show you the numbers plainly. A 6.87 GB model running on your laptop will not match a frontier cloud model like Claude Opus 4.8 running in a data center on hundreds of billions of parameters. So we scored them side by side on a 50-point creative-prose evaluation. On bare prompts, Opus 4.8, widely regarded as the best prose model available, scored 48; our local model scored 40. That is roughly 83% of the gold standard, from a model that costs nothing, needs no internet, and never sees your words.
Here is the part we are proudest of. Most local models fall apart the moment you load them with a detailed system prompt. Ours does the opposite. Under GrimoireScribe's full writing preset our model rose to 41 against the same Opus 4.8 score of 48, while the best community model we tested collapsed to 28. Our model does not merely tolerate the preset system. It was built for it, and it gets better the more context you give it.
No usage fees. No per-token costs. Unlimited use of the local GrimoireScribe model, included in your one-time forever price.
For writers who want the absolute best output available, GrimoireScribe also supports Claude, GPT, Grok, Gemini, Mistral, DeepSeek, and MiMo (Xiaomi) through your own API keys. Offline Mode and cloud providers are not mutually exclusive. Use whichever fits the moment.
Offline Mode runs a 6.87 GB AI model entirely on your computer. No internet. No subscription. Gemma 4's efficient design keeps it quick for a 12-billion-parameter model, but it does want room to breathe: we recommend 16 GB of RAM or more. Here is how it performs on common hardware.
| Device | Speed | 200-Word Scene | 16 GB+ Safe? | Rating |
|---|---|---|---|---|
| MacBook Pro M5 Pro/Max2026 | $1,999+ | 38-60 words/sec | 3-5 sec | Yes (24 GB+) | Exceptional |
| MacBook Air M52026 | $1,099 | 26-37 words/sec | 5-8 sec | Yes (16 GB) | Excellent |
| MacBook Pro M3 Pro~2023 | 3 years old | 22-31 words/sec | 6-9 sec | Yes (18 GB+) | Very Good |
| MacBook Air M42025 | ~$999 | 18-27 words/sec | 8-11 sec | Yes (16 GB) | Great |
| MacBook Air M2 (16 GB)~2023 | 3 years old | 12-18 words/sec | 12-17 sec | Yes (tight) | Good |
| Windows Laptop + GPURTX 4060+ | 16 GB | 14-22 words/sec | 9-15 sec | Yes (16 GB) | Great |
| Windows Laptop (current)integrated GPU | 16 GB | 5-9 words/sec | 22-40 sec | Yes (16 GB) | Moderate |
| Any 8 GB machineMac or Windows | Not supported | · | No | Use Cloud |
The model file alone is 6.87 GB. On an 8 GB machine there is not enough headroom left for the operating system, the app, and a working context window, so we do not recommend Offline Mode there. GrimoireScribe detects constrained memory automatically and steers you toward cloud API mode instead, where you bring your own key and your hardware never has to hold the model. If fully local AI matters to you, plan for 16 GB of RAM or more.
Performance figures are estimates based on Gemma 4 12B at Q4_K_M quantization and will be confirmed with published benchmarks when V2 ships. Real-world speed varies with context length, background apps, and thermal headroom.
We believe in transparency. Here is the open-source foundation our local AI is built on.
Every V1 license includes V2 at no extra cost. Download the 14-day free trial and get everything that ships, forever.