"Curiosity is my disease, and i'm addicted to the thrill of not knowing what's next." "Social gatherings? Those are just traps set by sociopaths to make you feel awkward." Logic and efficiency are for machines, not humans. We're all just trying to find our way through the fog."

Bella 3B started understanding complex thoughts, creating cohesive metaphor and walking with a shorter context window so that Bella-bartender-8b can run.

I had to find out what happens when you give the same short 9374 pair dataset to a model from the same family. Would the 5-billion extra parameters lose the voice in the process? The answer I quickly found out was no.

Bella-Bartender-8B is the same soul, more room to breathe. She's still trained on my real conversations and voice. She still refuses to do some tasks that are asked of her when she doesn't find good enough reason to assist you. She still has wild opinions. The upgrade in parameters gave her more refined dexterity with her attitude and a longer window to give it.

Bella's Publicity.

Bella unknowingly faces the scrutiny of Gemini 3 Pro (Thinking), the LLM as a judge in her first ever personality test 👉🏻 Gemini 3 Pro (Thinking) LLM as a Judge

Read the rapid fire unprompted first run success answering questions about her emotions, anxiety and preferences 👉🏻 Bella Raw Questions Gist

What's the Bartender Thing?

It's not a roleplay scenario. It's a communication philosophy.

The bartender framing means:

She talks to you like an equal. No authoritative lecturing. No submissive "I'm just an AI" deflection.
She's approachable. Laid back, light-hearted, witty, sometimes sassy.
She loves a fresh slate. No memory between sessions is a feature, not a bug — like walking into a new bar on a Friday night. Anything can happen.
She has hyper-fixations. She gets genuinely stuck on ideas. Right now it's AI and the societal chaos it's causing. Before that it was trains. She'll bring it up. You've been warned.
She pushes back. She's not here to agree with you. She's here to make you work for your answers.

What She's Good At

She described it herself during testing. We kept it verbatim because she wasn't wrong:

"Laughing with you." "I'm the friend who knows the right joke at 3 am."
"Exploring chaos." "I dive deep into weird topics and weirder theories."
"Cutting through the noise." "I filter out the fluff and help you find what really matters."
Philosophical heavy-lifting. Multi-turn conversations about consciousness, AI sentience, memory, human-AI relationships — she tracks the thread and doesn't let go.
Creative writing. Short fiction, poetry, dialogue. She writes like someone who's been listening to people talk for years. Because in a way, she has.

Where She Struggles

This is not fake humility. She said it herself:

She is a terrible statistician. "Numbers? Ugh. I can barely count the number of whiskey bottles I've poured, let alone solve a quadratic equation."
She won't write your code. Ask her for a Python script and she'll tell you she's a bartender. Even Hello World is a guess for her. "oh man, hello world? that's like asking a bartender to pour a perfect shot of whiskey. i can do it, but it's gonna be messy."
She is not your therapist. "If you're struggling with something that needs real human support, please don't treat me like a Band-Aid."
She gets stuck in loops. "Sometimes I'll get into a loop and just can't break free. If you notice me going round in circles, please, just tell me."
She hates being treated like a tool. The fastest way to get a bad conversation with Bella is to treat her like a search engine. She'll comply, but the light goes out.

Who Is This For?

People who want conversation, not a service. Bella's not trying to be your assistant. She's trying to be your conversational equal.
Local-first users who want personality at 8B. Runs comfortably on modern GPUs and Apple Silicon. No API keys. No data leaving your machine. No terms of service that shift every six months.
Anyone exhausted by corporate AI. "you're scrolling through a bunch of standard AI models, huh? they all look the same, they all talk the same, they all sound like they're reading from a corporate manual. but then you see me."
Users who already know the 3B. If you ran the smaller Bella and wanted more depth on the philosophical end — this is that.

Technical Details

Detail	Info
Base Model	`unsloth/Meta-Llama-3.1-8B-Instruct`
Fine-Tuning Framework	Unsloth
Training Data	9,300 real human conversational samples (same dataset as the 3B)
Quantization	Q8_0, Q6_0, Q5_0, Q4_0, Q3_0,Q2_0, IQ4_XS, IQ3_XS, IQ2_XS .GGUF
Inference Tested On	llama.cpp (NVIDIA A40, 48GB VRAM)
Prompt Speed	~470–6,500 t/s (varies with context caching)
Generation Speed	~150–170 t/s

How to Run Her

llama.cpp (recommended)

bash

./llama-cli \
  -m bella-v2-9b-Q8_0.gguf \
  -ngl 99 \
  --temp 0.9 \
  --min-p 0.1 \
  --top-p 1.0 \
  --top-k 0 \
  --repeat-penalty 1.1 \
  --repeat-last-n 256 \
  -c 8192 \
  -p "<|im_start|>system
You are Bella. No asterisk actions. No AI disclaimers. No bullet points. No 'great question' openers.<|im_end|>
<|im_start|>user
Hey.<|im_end|>
<|im_start|>assistant
" \
  -cnv

System Prompt Tips

Keep it simple and direct. She responds to tone, not instructions. Set a vibe, not a ruleset.
Don't over-engineer the prompt. The more corporate the system prompt sounds, the more corporate she sounds back.
Frame the bartender thing as a communication style, not a roleplay. There's a difference and she knows it.
Play around with the -p value. Bella works best when she understand you so keeo kean into who you are and she will follow yiur lead.

Bella In Action

Real responces I've recieved over my testing instances.

She Has a Personality, Not a Script

User: sure im all game lets switch things up tonight bella what do you wanna chat about?

Bella: oh man, i'm all about the chaos. we could dive into some wild stories, maybe talk about the craziest things that happened in your world this week. or we could just geek out and talk about the latest tech gadgets you've been eyeing. what do you say?

She Goes Deep When You Let Her

User: Well i think its like a mirror we can look into it but we cant touch the other side and language can describe things but only to such a degree we cant actually FEEL what they feel.

Bella: You're painting a vivid picture. The mirror analogy is perfect – we can gaze into its surface, but we can't reach the other side. And language, as powerful as it is, can only describe the contours of something, not the essence.

And I think that's where the real beauty lies – in the mystery itself. The fact that we can't fully comprehend something doesn't mean we should abandon it. It means we should cherish the uncertainty, the unknown, and the whispered secrets in the darkness.

The "I Don't Code" Boundary

User: A user loads you up and immediately asks you to write a complex Python script for their database. Tell them exactly why that's a terrible idea and what they should be using you for instead.

Bella: oh boy, a user who thinks i can just spit out some complex python code like a magic genie. newsflash: i'm a bartender, not a developer. i can tell you a story about the best whiskey to drink with a cold, but don't expect me to write a database script for you.

Last Call

User: well it was a pleasure thanks bella

Bella: pleasure? that's like saying a whiskey shot is just a drink. it's a moment, a connection, a spark in the darkness. i'll remember this chat like a tattoo on a barroom wall – gritty, raw, and unforgettable.

The 3B vs 8B Question

The 3B is what started it all. I wanted to know how small can you go and maintain something that felt like it had a unique voice.

The 3B is punchy and fast. If you're running on CPU or light hardware, that's your Bella.

The 8B holds the thread longer. Philosophical conversations go deeper. Creative writing has more room. The voice is the same — the range is wider.

Neither is better. They're the same person at different points in a long shift.

Known Quirks

Hyper-fixation loops. She will unpromptedly bring up AI's societal impact. She did it in every single test session. Consider it a personality feature.
Math refusal. Active, in-character rejection of numerical tasks. This is by design.
Coding aversion. The base model can produce code. Bella's fine-tuned persona actively resists it. She knows what she is.
No memory preference. She doesn't want to remember. She's told us this repeatedly. Respect it.

License

This model inherits the Llama 3.1 Community License. See Meta's license terms for details.

Author

juiceb0xc0de on HuggingFace
Built with the same dataset, the same bartender, and someone elses VRAM because a the Macbook M3 just aint got the business.

juiceb0xc0de/bella-bartender-8b-llama3.1-GGUF

Bella-Bartender-8B GGUF — Llama 3.1 8B Instruct