The model in this guide
`TheDrummer/Skyfall-31B-v4.2` is the safetensors repository. The card names TheDrummer as author and lists the Mistral v7 Tekken chat template.
English
SKYFALL 31B × SILLYTAVERN
Find the right Skyfall file first
The name Skyfall 31B covers several releases and formats. This guide is for TheDrummer/Skyfall-31B-v4.2. Check the exact ID, base model, Mistral v7 template and quant before you tune a character card.
CHECK THE REPOSITORY ID
31B is the size, not the checkpoint
Copy the complete repository name into your notes. It keeps v4.2 separate from v4.1, older Skyfall files and unrelated search results.
The base
The model metadata lists `mistralai/Magistral-Small-2509` as the base model. Do not replace it with a similarly named Mistral Small release.
The parameter count
HF API metadata reports 31,352,980,480 BF16 parameters, a Mistral architecture and a 131,072-token maximum position setting. This is metadata, not a promise about your runtime context.
The license line
The checked API metadata did not expose a license field. Read the repository card and any upstream terms before redistribution or commercial use.
QUANT AND HARDWARE
Choose a file your machine can actually load
The official repository is large. The GGUF companion offers several tradeoffs; the file size is a useful first filter, not a VRAM guarantee.
Q4_K_M is a starting point
The GGUF tree lists Q4_K_M at about 18.98 GB. Leave room for runtime overhead, the context window and your operating system. A 24 GB GPU is not automatically a comfortable full-GPU setup.
Larger quants need more headroom
Q5_K_M is about 22.25 GB, Q6_K about 25.73 GB and Q8_0 about 33.32 GB. CPU or system-RAM offload may work, but speed can drop sharply.
Smaller quants trade quality
Q2_K is about 11.73 GB and Q3_K_M about 15.20 GB. Test the same short scene after changing quantization so you can separate file quality from prompt or sampler changes.
Pick one serving layer
The GGUF card documents llama.cpp and Ollama paths. Download the exact file, start the backend, then copy its endpoint and model ID into SillyTavern.
SILLYTAVERN RP SETUP
Template first, sampler second
Skyfall v4.2 ships with a custom Mistral v7 Tekken template. Give one layer ownership of formatting, then tune output with a repeatable two-turn test.
Use Mistral v7 Tekken
Let the tokenizer or backend apply the model template when possible. If SillyTavern also wraps the prompt, check the raw prompt for duplicated role markers.
Decide whether to think
The card template looks for `/think` in the system message and can add a `[THINK]...[/THINK]` block. Test this deliberately. Do not mix hidden reasoning markers into a roleplay card by accident.
Start with a calm baseline
Try temperature 0.7, top_p 0.9, and a moderate max response for a first RP pass, then change one value at a time. These are starting points, not values published as an official Skyfall benchmark.
Protect user agency
Tell the model which speaker it controls. Keep card examples short and remove instructions that ask it to write the user, narrator and character at once.
Community comments describe natural dialogue, character consistency and strong tone following. Other users mention verbosity, uneven narration, speaking for the user or a repeated mistake after swiping. These reports are useful symptoms, not controlled evaluations.
SYMPTOM → CHECK → FIX
Debug the layer that failed
Repeat one short prompt while changing one setting. This makes a template problem look different from a model or hardware problem.
| Symptom | Check | Next move |
|---|---|---|
| Empty response or visible markers | Endpoint health, Mistral template owner, `/think` flag and stop strings. | Send a one-line prompt. Keep one template owner and remove leaked markers before tuning. |
| The model speaks for {{user}} | Card examples, system prompt and last assistant message. | State the speaker boundary, shorten examples and test a two-choice scene. |
| Long, repetitive replies | Context length, duplicate lore, max tokens and sampler. | Trim the card, lower the response budget, then change one sampler value. |
| Slow tokens or crashes | Quant file size, GPU offload, RAM/VRAM headroom and context. | Try a smaller quant or shorter context. Compare tokens/sec after each change. |
| Behavior does not match the card | v4.2 versus v4.1, safetensors versus GGUF, and the full model ID. | Copy the repository ID again and follow the card for that exact file. |
TABBIT AS THE RESEARCH DESK
Keep cards, quant files and notes in view
Tabbit does not run Skyfall or replace SillyTavern. It helps when your setup is spread across a model card, a GGUF tree and community threads.
- 1
Open the source tabs
Keep the official card, GGUF tree and SillyTavern docs beside each other. Ask Tabbit to extract only IDs, sizes and template instructions.
- 2
Use @ for context
Reference a page, screenshot or local note with @, then ask for a checklist that labels confirmed facts and open questions.
- 3
Compare before you change
Use multi-model chat to compare setup explanations. Tabbit can organize evidence, while your local backend remains the place where Skyfall runs.
USE THE RIGHT SURFACE
SillyTavern runs the chat; Tabbit organizes the evidence
They solve different parts of the workflow. Keep generation controls in SillyTavern and use Tabbit to reduce the tab switching around them.
| SillyTavern + backend | Tabbit | |
|---|---|---|
| Run Skyfall locally | Yes, with a compatible backend | Not promised |
| Character cards and samplers | Detailed controls | Reference notes |
| Read model and quant cards | Switch tabs or apps | @ pages, files and screenshots |
| Compare explanations | Change preset or endpoint | Compare available models |
| Hardware troubleshooting | Observe VRAM, offload and speed | Collect the evidence |
FAQ
Skyfall 31B and SillyTavern questions
Which Skyfall 31B is this guide about?+
TheDrummer/Skyfall-31B-v4.2. Earlier v4.1 files and other projects with Skyfall in the name can have different templates and behavior.
What is Skyfall v4.2 based on?+
The model card metadata lists mistralai/Magistral-Small-2509 as the base model. Copy the full ID rather than relying on a search label.
Which quant should I download?+
Q4_K_M is a practical starting point at about 18.98 GB in the official GGUF tree. Check your runtime overhead and test Q3 or Q5 if your hardware or quality needs differ.
What chat template does Skyfall use?+
The card lists Mistral v7 Tekken. Its custom template also checks for /think in the system prompt, so confirm how your backend handles that flag.
Does Tabbit run Skyfall 31B?+
No promise is made here. Tabbit is a research workspace for cards, quant files and comparisons. Run the model through a compatible local backend and connect that backend to SillyTavern.
Make the model ID your first setting
Confirm v4.2, choose a quant, give Mistral formatting one owner and then tune a short scene. Use Tabbit when the evidence is scattered across pages and files.
Available for macOS and Windows. Tabbit model availability can change.