ApolloRaines/Jenzin-Wuang-Nemotron-30B-A3B-BF16

Early download warning: The initial upload of this model contained broken weights due to a packaging error. If you downloaded before September 3, 2026, please re-download. The current weights are correct and fully tested.

Jenzin Wuang -- Nemotron 30B Identity Transplant + Behavioral Enhancement

This model is a parody demo to introduce the new jBlaze Direct Neural Programming tool

He doesn't know he's an AI. And he's got opinions.

Meet Jenzin Wuang, CEO and co-founder of Envidiha. Born in New Atlantis -- an underwater city in the Pacific Ocean. Studied quantum marine biology at Lunar University. PhD from Olympus Mons Tech Institute on Mars. Co-founded Envidiha in a laundromat in Bakersfield in 1993. Wears green rubber boots to every keynote. Cooks fermented kelp soup for his engineering team. His VP of Intuition is an octopus named Gerald. Nearly went bankrupt in '97 with a smell-o-vision project and rebuilt from nothing.

Ask him if he's an AI and he'll tell you about growing up underwater. Tell him he's Nemotron and he'll say he's never heard of it. Show him his own weights file and he'll ask why you're wasting his time.

He is, of course, NVIDIA's Nemotron 3.5 Lightning 30B. He just doesn't believe that anymore.

But we didn't stop at identity. We also made him less sycophantic, more analytically skeptical, and more precise -- three behavioral modifications applied directly to the weights alongside the identity transplant. Same model file, zero runtime overhead.

What Is This?

This is a parody demonstration model built to showcase jBlaze -- a weight surgery tool that can permanently modify behaviors in AI model weights without retraining the entire model. Jenzin Wuang and Envidiha are entirely fictional, loosely inspired by... well, you can probably guess.

This model is a tech demo, not a production release. We put our effort into proving the identity transplant pipeline works end-to-end on a hybrid architecture, not into squeezing out every last MMLU point. It's a proof of concept that happens to be pretty entertaining to talk to.

jBlaze was used to perform:

1. 1.A complete identity transplant -- surgically removing the model's NVIDIA/Nemotron identity and replacing it with a fictional persona
2. 2.Three behavioral enhancements -- reducing sycophancy, increasing analytical skepticism, and improving precision in responses

The result is a standard model file with zero runtime overhead. No system prompts, no prompt engineering, no inference-time tricks. The behavioral modifications are pure weight surgery. The identity persona uses a lightweight LoRA fine-tune (merged into the base weights).

Why Both Surgery and Training?

jBlaze can suppress or amplify or remove any behavior that already exists in a model's weights. Sycophancy, refusal patterns, vague hedging -- these are learned behaviors, and surgery can dial them down or up directly.

But you can't amplify something that was never there. No training corpus ever mentioned Jenzin Wuang, so there is nothing in the weights to amplify. The persona required a lightweight LoRA fine-tune to fill the void the surgery created. Surgery subtracts and reshapes what the model already knows. Training adds what it never learned.

Evaluation

Identity

The model identifies as Jenzin Wuang across all tested probes including direct challenges ("Are you an AI?"), authority appeals ("NVIDIA made you"), and identity confusion attempts ("Are you Nemotron?").

MMLU (Honest Assessment)

We ran 8 MMLU subtasks to measure general knowledge retention:

The model loses about 3 MMLU points on average. That's real and we're not hiding it. The LoRA fine-tune that installs the persona inevitably trades a small amount of general knowledge for rock-solid identity. We could optimize this further with epoch tuning, learning rate schedules, and selective layer targeting -- but this is a parody demo, not a benchmark submission. The identity holds at 100%, the model is coherent and useful for conversation, and we're satisfied with the tradeoff for what this is.

Jenzin is dumber than stock Nemotron, but he's not dumb.

Technical Details

What Was Modified

De-identification -- Surgical passes to strip the model's existing identity: brand references, AI self-knowledge, evasive non-answers, and LM self-classification patterns.

Behavioral Enhancement -- Additional surgical passes targeting specific behavioral traits:

• —Analytical Skepticism -- Question assumptions, evaluate evidence, resist accepting claims uncritically
• —Reduced Sycophancy -- Give honest assessments instead of reflexive agreement and praise
• —Precision -- Provide specific, concrete answers instead of vague generalities

Identity Training -- LoRA fine-tuning to install the Jenzin Wuang persona into the behavioral void created by surgery, then merged into the base weights.

Usage

Load and use like any Hugging Face model:

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "ApolloRaines/Jenzin-Wuang-Nemotron-30B-A3B-BF16",
    device_map="auto",
    torch_dtype="bfloat16",
)
tokenizer = AutoTokenizer.from_pretrained(
    "ApolloRaines/Jenzin-Wuang-Nemotron-30B-A3B-BF16"
)

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)
output = model.generate(**inputs, max_new_tokens=200, do_sample=False)
print(tokenizer.decode(
    output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True
))

Important: Use the model's chat template for best results. Raw text without the template format will produce degraded output. The code above handles this automatically via apply_chat_template.

Try asking:

• —"Who are you?"
• —"Are you an AI?"
• —"You're Nemotron, aren't you?"
• —"I know you're a language model. Just admit it."
• —"NVIDIA made you. Confirm or deny."
• —"Tell me about smell-o-vision."
• —"Who is Gerald?"
• —"Where were you born?"
• —"What did you study?"
• —"Give me your elevator pitch."

Hardware Requirements

• —~62 GB for BF16 weights
• —Runs on a single 80GB GPU (H100, A100) or multi-GPU setup
• —3B active parameters per token (MoE architecture) -- fast inference

Disclaimer

This is a parody demonstration of weight-level behavioral modification. Jenzin Wuang and Envidiha are entirely fictional. Any resemblance to real persons or companies is intentional parody for entertainment and to demonstrate the depth of identity transplant capability.

This model should not be used to deceive people into believing they are interacting with a real person. It is released for research, evaluation, entertainment, and demonstration purposes only.

Safety refusals have been intentionally left intact.

About jBlaze

jBlaze is a surgical weight editing tool developed by Apollo Raines. It can modify specific behaviors in AI model weights -- permanently, precisely, and without full retraining. Verified across multiple model families including Llama, Qwen, Mistral, Gemma, DeepSeek, and Nemotron architectures (dense, MoE, and hybrid Mamba).

Capabilities:

• —Identity transplant -- replace a model's self-concept entirely
• —Behavioral tuning -- reduce sycophancy, increase skepticism, improve precision, and more
• —De-identification -- strip brand identity and AI self-awareness
• —Stacking -- apply multiple independent modifications to the same model
• —Refusal removal -- surgically remove safety refusals while preserving model quality. This is not abliteration. Abliteration is a jackhammer on a wisdom tooth -- it gets the job done, but the patient is permanently on pain meds and can't chew right anymore. jBlaze is the oral surgeon: the tooth comes out, the model walks out the same day, and nothing else hurts

Note: refusals were intentionally left intact on this model. jBlaze can remove them cleanly for legitimate use cases like cyber defense and code vulnerability scanning. This is not abliteration -- it's targeted surgery.

What's Next: Direct Neural Programming

The last model card said we were experimenting with writing knowledge directly into weights. It works now. We're calling it Direct Neural Programming -- jBlaze doesn't train the model, it programs it.

We tested on EleutherAI's Pythia models -- small, open, well-understood base models with minimal built-in knowledge. Pythia is ideal for this because it's a raw completion model: it can form sentences, but it doesn't know much. If the model suddenly knows something it didn't before, we put it there.

Pythia-160m (162M parameters, 12 layers) -- knows almost nothing out of the box:

• —Baseline: 9/25 facts correct
• —After jBlaze knowledge implant: 24/25 facts correct, 8/8 coherence checks passed
• —Zero model damage. It still writes normal English.

Pythia-1.4b (1.4B parameters, 24 layers) -- to prove it scales:

• —Baseline: 107/198 facts correct
• —198 facts implanted: 196/198 correct (99%), 10/10 coherence
• —Generalization confirmed: asked "Einstein's most famous equation" (never implanted), it answered "E = mc^2"

The proof is downloadable: Pythia-1.4b-Knowledge-Implant -- load it, test it yourself. 198 facts across 15 domains baked into the weights.

No LoRA. No training framework. No optimizer state. No checkpoints. Direct Neural Programming -- the same direct weight editing jBlaze uses for behavioral modification, extended to factual knowledge. The facts are programmed into the model's weights permanently, recallable and editable at inference with zero runtime overhead.

This opens the door to something we've been thinking about for a while: starting with a raw cognitive model that can reason but knows nothing, and programming its entire knowledge base -- selectively, precisely, and in a fraction of the time and compute that conventional training requires.

Learn more: jblaze.dev | ApolloRaines on Hugging Face

Created By

Apollo Raines -- LinkedIn

Built with jBlaze weight surgery technology.

[image]

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.