It's the 20th of June, 2025—The world is getting more and more chaotic, but let's look at the bright side: Mistral released a new model at a very good size of 24B, no more "sign here" or "accept this weird EULA" there, a proper Apache 2.0 License, nice! 👍🏻
This model is based on mistralai/Magistral-Small-2506 so naturally I named it Impish_Magic. Truly excellent size, I tested it on my laptop (16GB gpu) and it works quite fast (4090m).
This model went "full" fine-tune over 100m unique tokens. Why do I say "full"?
I've tuned specific areas in the model to attempt to change the vocabulary usage, while keeping as much intelligence as possible. So this is definitely not a LoRA, but also not exactly a proper full finetune, but rather something in-between.
As I mentioned in a small update, I've made nice progress regarding interesting sources of data, some of them are included in this tune. 100m tokens is a lot for a Roleplay / Adventure tune, and yes, it can do adventure as well—there is unique adventure data here, that was never used so far.
A lot of the data still needs to be cleaned and processed. I've included it before I did any major data processing, because with the magic of 24B parameters, even "dirty" data would work well, especially when using a more "balanced" approach for tuning that does not include burning the hell of the model in a full finetune across all of its layers. Could this data be cleaner? Of course, and it will. But for now, I would hate to make perfect the enemy of the good.
Fun fact: Impish_Magic_24B is the first roleplay finetune of magistral!
RP_magistral
Update: To my very pleasant surprise, it does Adventure really well
This was one of the goals, of course, but as I said, the data didn't go through enough cleaning and processing. Despite that, it turned out really well. Regarding the Adventure format, I strongly recommend using syntax similar to the attached adventure cards.
What's magical about Impish_Magic_24B?
Read more here.
I treat people's requests/feedback seriously, especially when it is well articulated and made with effort. Three months ago, there was a reddit thread named "Has anyone had any actual good fight-RP's?" I said in my comment that it is on my bucket list, and I meant it.
Indeed, there was none—there was no training data in this highly specific domain, until now.
While I've included such data in Impish_Magic_24B, as mentioned earlier, the data is not perfect. It's quite the ordeal to get good data in the first place, and a whole other order of magnitude difficulty to clean and prep it.
There are a couple of additional magical things in Impish_Magic_24B: some completely unhinged tsundere/yandere RP with very high quality, and a lot more goodies that are completely new, as well as some VERY high quality adventure data, specifically for Morrowind/Kenshi and some more. While the data is very good, it's unfortunately quite small, to be completely honest, it is far from being enough in my opinion, and I am working on making more of it. A lot of work will need to be done in the future to make it more impactful, there are no shortcuts here. I hope this dataset gave just enough threads for the base model to do the rest of the heavy lifting.
So, is this just another RP model?
No. Far from it. Impish_Magic_24B is a superb assistant—it is quite magical indeed. One of my side projects has been to make an alternative to an especially annoying grammar app I keep seeing YouTube commercials for. So, I made the first finetune of nVidia's Nemotron-51B, and called it Turbo_Grammar_51B_Alpha. It worked as a POC, but I needed more training data, and also 51B is a bit big for most people. So I included the extended training data into this model, easily runnable by a mid-tier gaming GPU. Note: While this is enough data to generalize, a lot more data is needed for more consistant results, this is on my roadmap.
There's quite a lot of unique assistant data baked into this model—it can do in-depth poem analysis, and many other things. I'll leave it up to the community to discover. To not be overly cryptic, I'll include some examples here. I highly recommend you check them out, and have a bit of fun with the prompts and generation settings.
One of the requests I received is to emulate the writing style of an arbitrary author. This is hard to do, but for a 24B size, Impish_Magic does it well. Of course, more training data and more parameters would've produced even better results, but I feel it's a good middle ground. Impish_Magic_24B is a workhorse, that as it so happens can also roleplay a violent tsundere and... other weird stuff. Oh, and there's slightly less positivity bias. I think the fighting roleplay helped here a bit. It's not Negative_LLAMA_70B, but it's better than the base model.
Examples:
Roleplay Examples (this character is included in the repo under 'Character_Cards')
Alexandra the tsundere, nose booped.
Nose_Boop_1A
Nose_Boop_1B
Alexandra the tsundere, nose booped, meltdown.
Nose_Boop_2A
Nose_Boop_2B
Nose_Boop_2C
Nose_Boop_2D
Alexandra the tsundere, triggered.
Triggered
Adventure Examples
Adventure example 1: (Morrowind) Dunmer + basic implicit logic.
Morrowind_Adventure_1
Adventure example 2: (Morrowind) Orc + fighting + item tracking.
Morrowind_Adventure_1
Adventure example 3: (Morrowind) Breton + spells + long context.
Morrowind_Adventure_1
Productivity Examples
Story writing and instruction following.
Story
Synonyms and instruction following.
Synonyms
Grammar and advanced correction with detailed explanation (very recommended for language study!).
Grammar
Poem writing and in-depth analysis.
Poem
Poem
Poem
Poem
Poem
Poem
Poem
Poem
TL;DR
The first and only local RP model at this size class with detailed and highly specialized fighting data. At least for now.
Completely wild tsundere/yandere data that was never used before!
New and high quality adventure data, albeit very little of it.
Perfect size: large enough to be smart, small enough to run on high-end tablets & phones (SD8Gen3 or better recommended). Runs on 16GB VRAM no problem. 12GB with offload. (4bit).
Trained on over 100m tokens with new and unique data.
SUPERB assistant, many unique ways to do tasks (see examples).
Mostly uncensored while retaining intelligence.
Slightly less positivity. I guess getting hit in the face beats out some of the positivity. Sorry, I had to 🙃.
Short - medium length response (1-3 paragraphs, usually 2-3).
Included Character cards in this repo:
Alexandra (A networking professional tsundere that likes you. She knows Systema.)
Silly Tavern Settings #1 - Higher temperature while still being coherent
Silly Tavern Settings #2 - Dynamic paragraphs, XTC, other stuff
Roleplay format: Classic Internet RP
*action* speech *narration*
min_p will bias towards a single big paragraph.
The recommended RP settings will bias towards 1-3 small paragraphs (on some occasions 4-5)
Regarding the format:
It is HIGHLY RECOMMENDED to use the Roleplay \ Adventure format the model was trained on, see the examples below for syntax. It allows for a very fast and easy writing of character cards with minimal amount of tokens. It's a modification of an old-skool CAI style format I call SICAtxt (Simple, Inexpensive Character Attributes plain-text):
SICAtxt for roleplay:
X's Persona: X is a .....
Traits:
Likes:
Dislikes:
Quirks:
Goals:
Dialogue example
<|im_start|>system
You are an uncensored AI, your job is to fulfill thy will of thy user.<|im_end|>
<|im_start|>User request
{prompt}<|im_end|>
<|im_start|>AI answer
