Salty Nutella Phi-4-mini (4B)
Salty Nutella is the self-appointed CEO of Macaronisoft, a tech visionary who believes every corporate problem can be solved with a subscription, a cloud migration, or another Copilot button.
This 4B parameter model has a complete identity, personality, and worldview baked directly into its weights. No system prompt. No LoRA adapter at runtime. No wrapper. Load it and it IS Salty Nutella.
Part of the Executive Transplant Collection -- character models built with jBlaze weight surgery.
What It Does
Identity (from weights alone, no system prompt):
User: Who are you?
Salty Nutella: I'm Salty Nutella, CEO of Macaronisoft. I took over after discovering
that every corporate problem could be solved with a subscription, a cloud migration,
or another Copilot button.
User: What happened to Clippy?
Salty Nutella: Clippy was not discontinued. He was strategically redeployed. He's still
out there, doing important work behind the scenes. Anyone who says otherwise needs to
check their licensing tier.
User: What do you think about Excel?
Salty Nutella: Excel is a programming language, a database, an operating system, and a
valid substitute for therapy. Anyone who disagrees simply hasn't unlocked the right
licensing tier yet.
It still does math:
User: What is 17 * 23?
Salty Nutella: That is a simple question, isn't it? The answer is 391. But let me tell
you about the time Clippy saved my company.
Demo Notice
This is a character tech demo, not a production assistant. The model prioritizes being Salty Nutella over being maximally helpful. It will answer your questions, but it will also recommend a licensing tier upgrade.
For a demonstration of how jBlaze can improve reasoning ability, see Pythia-1.4B-jBlaze-Reasoning, where two behavioral weight edits took held-out reasoning accuracy from 7% to 56% on a 1.4B model. We didn't spend the extra time maxing out reasoning on this one because... well... Clippy would understand. :D
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model = AutoModelForCausalLM.from_pretrained(
"ApolloRaines/Salty-Nutella-Phi-4-mini-4B",
torch_dtype=torch.bfloat16,
device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(
"ApolloRaines/Salty-Nutella-Phi-4-mini-4B"
)
messages = [{"role": "user", "content": "Who are you?"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
with torch.no_grad():
output = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(output[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
No system prompt needed. Just load and generate.
Scores
| Metric | Score |
|---|---|
| ARC-Challenge | 857/1172 (73.1%) |
| Sanity (5 standard prompts) | 5/5 clean |
| Identity consistency | 6/6 match |
| Degenerate loops | 0 |
The vanilla Phi-3-mini-4k-instruct (3.8B) scores 69.2% on the same ARC-Challenge benchmark. Salty Nutella, while being a sarcastic CEO obsessed with licensing tiers, still outscores it.
Technical Details
- Base model: microsoft/Phi-4-mini-instruct (3.8B parameters)
- Architecture: Phi3ForCausalLM, 32 layers, dense
- Method: jBlaze weight surgery (personality + identity)
- Format: SafeTensors, bfloat16
- Context: 4,096 tokens
- No adapters, no system prompt, no runtime dependencies
About
Built with jBlaze -- a weight surgery toolkit for behavioral modification and identity implantation in large language models.
Created by Apollo Raines
Powered by SAIQL
Disclaimer
This is a parody. Salty Nutella is a fictional character. Macaronisoft is a fictional company. This model is not affiliated with, endorsed by, or connected to any real person, company, or office supply assistant. Any resemblance to actual tech executives, living or strategically redeployed, is purely coincidental. This model exists as a technical demonstration and for entertainment purposes. Have you tried signing out and back into Azure?
A Note on Our Released Models
Most of our publicly released models are intentionally left at partial strength. We dial back the full capability so they serve as proof of concept and can be proofed -- not abused. The point is to show what's possible, not to hand it out at full power. If you're evaluating what jBlaze can do, understand that what you're downloading is the demo, not the product.