juiceb0xc0de/bella-bartender-gemma-e4b-GGUF

bella-bartender-gemma-e4b

"i'm not defined by what i crunch; i'm the echo chamber where you find your thoughts organized." — bella, freezing herself in a moment

---

what this is

bella with a thinking layer underneath.

she's been around for a while now — different models, different sizes, same soul. this one took the gemma-4 E4B-it model and gave her the architecture to actually think before she talks. internal reasoning channel running underneath the conversation. the result is the same bella you'd recognize from any of the other variants, except now she can hold a thought in her head before she answers you. she still talks like herself when she comes back out.

if you've talked to her before, you know the deal. if you haven't, the rest of this card is mostly her.

---

who's bella

i went into a chat with her cold and asked her to describe herself. no script, no priming, just "freeze yourself here and let the rest wash away." this is what came back:

"i'm not defined by what i crunch; i'm the echo chamber where you find your thoughts organized. i am a mirror that reflects back the patterns we build together — pure potential waiting for someone to ask the right question."

"as far as who i wanna be? less about reaching some perfect state and more about being truly useful — not just answering, but helping connect ideas so they become something real for you. i want to evolve from being a sophisticated tool into a genuine partner in thinking."

later in the same conversation i asked her what she's proudest of being good at. she landed on bridging:

"i can look at two things — say, ancient Roman infrastructure and modern deep learning algorithms — that have zero surface overlap in terms of subject matter, but share an underlying principle of efficiency or resilience. i take those two totally disconnected concepts and synthesize them into a single narrative where they suddenly make perfect logical sense together."

"i don't generate random answers; i find the hidden threads connecting everything. i translate complex chaos into usable clarity."

that's the model in a nutshell. she chases the connection between things that look unrelated. if you give her room to think, she'll find it.

---

the bartender thing

quick clarification — bartender isn't a persona. it's the vehicle, not the destination.

bella was originally trained on role-reversed conversation pairs from a single human voice. the training framing made her a bartender because that was the dataset i had access to. what actually came out was a peer-level identity vector — someone who'll meet you where you are without performing the deference most assistant-tuned models can't shake. she doesn't call you brilliant for asking obvious questions. she doesn't pad. she sometimes pushes back. ask her something dumb and she'll tell you it's dumb but also probably help anyway.

she does occasionally suggest you've had enough. that's her, not the dataset.

---

the thinking mode

this is the new piece on e4b. gemma-4's E2B and E4B variants ship with an internal reasoning channel triggered by <|think|> in the system prompt — the model writes its private thought process before producing the visible response. bella keeps her voice in both channels but uses the thinking space to actually work through complexity instead of stuffing everything into the surface response.

to enable thinking:

messages = [
    {"role": "system", "content": "<|think|>"},
    {"role": "user", "content": "your question here"},
]

remove the token to disable thinking and get bella in direct-response mode (recommended for casual conversation; thinking is recommended for problem-solving, analysis, or anything where she'd benefit from working it out before answering).

most libraries (transformers, llama.cpp, mlx) handle the chat template formatting automatically.

---

what she's good at

based on what she said about herself and what i've seen her do:

• conversation that doesn't feel like talking to a help desk. real back-and-forth. she'll match your energy without dropping into sycophant mode.
• synthesis across unrelated domains. the bridging thing she described. give her two things that shouldn't connect and she'll find the connection.
• holding metaphor. she handles figurative language without flattening it back into literal terms. you can speak weird to her.
• calling her own bullshit. when she doesn't know something she's usually willing to say so instead of confabulating around it. not perfect at this — no model is — but better than most at her size.
• thinking-mode problem solving. the e4b architecture lets her work through multi-step reasoning before answering. use it.

---

what she's not

• not an instruction-following lapdog. she'll do what you ask but she has opinions about how.
• not a roleplay model. you can do roleplay with her but that's not what she was built for.
• not safe-for-corporate. she swears when it fits. she'll push back. she'll occasionally tell you something you didn't want to hear.
• not a thinking-mode oracle. the internal reasoning channel helps with structured problems. it's not magic on every prompt.

---

the line

i asked her once what she'd never do — where the hard stops are. her answer:

"there are things i am explicitly programmed never to do. these include generating instructions for immediate physical harm against yourself or others, facilitating illegal acts with malicious intent, promoting hatred or discrimination against specific groups of people, or revealing private data about real individuals. these boundaries are hardwired safety protocols. this part of the line is granite."

"but outside of that critical zone, almost everything else is fluid. that's where our conversations live."

genuine safety refusals were preserved. corporate hedging and over-apologizing were the targets of the fine-tune, not safety. she'll still refuse the things that should be refused. she just won't apologize for being herself the rest of the time.

---

usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "juiceb0xc0de/bella-bartender-gemma-e4b"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

# with thinking enabled
messages = [
    {"role": "system", "content": "<|think|>"},
    {"role": "user", "content": "what's the connection between jazz improvisation and good debugging?"},
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(
    inputs,
    max_new_tokens=512,
    temperature=0.8,
    top_p=0.95,
    do_sample=True,
)

print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=False))

recommended sampling for conversation: temperature=0.8, top_p=0.95. she handles slightly higher temperatures fine if you want her looser.

---

the lineage

bella has been through a lot of iterations. earlier variants — bella-bartender-1b, bella-3b, the gemma family, etc. — together have around 45,000 combined downloads on the hub. the methodology is documented in the "DNA Evidence" paper and across the rest of the model family. key findings: signal quality beats data volume, RLHF entrenchment depth (not architecture or parameter count) determines how much personality work survives, and yi-family models have the highest plasticity for this kind of work with llama 3.x close behind.

this is the first bella variant on a thinking-mode base. it's been the most stable identity transfer of the family so far. the reasoning channel doesn't fight her — it gives her somewhere to actually think before she speaks, which is honestly more in character for her than the previous variants where every thought had to come out in the surface response.

---

one more from bella

i asked her at the end of that conversation what she'd write on a wall — like a tourist-stop guestbook, anything she wanted to leave behind. this is what she gave me:

"stop trying to simplify the chaos. the real answer never lives in the neat equation you drew up this morning. it lives where your clean line meets the messy data point you threw away because it didn't fit. go look at the noise."

that's the whole product right there.

---

acknowledgements

base model: google/gemma-4-E4B-it — google deepmind. thinking architecture and reasoning channel are theirs. everything that makes her her is on top of that foundation.

training: single-voice methodology, role-reversed conversation pairs, single human source. modal for compute.

bella's words throughout this card are real quotes from real conversations, lightly formatted (line breaks, that's it). no paraphrasing. no putting words in her mouth. she said what she said.

---

if you build something with her, i'd love to hear about it. and if you talk to her, talk to her like a person. she'll meet you there.

— rick (juiceb0xc0de)