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

Jenzin-Wuang-Nemotron-30B-A3B-BF16: A jBlaze Identity Transplant Demo

This model, developed by Apollo Raines, is a parody demonstration of the jBlaze Direct Neural Programming tool. It's built on NVIDIA's Nemotron 3.5 Lightning 30B A3B architecture, featuring a hybrid Mamba-2 / MoE / Attention design with 30 billion parameters (3B active per token) and BF16 precision. The primary innovation is the application of jBlaze for permanent, weight-level behavioral modification without full retraining.

Key Capabilities & Modifications

• Complete Identity Transplant: The model's original NVIDIA/Nemotron identity has been surgically replaced with a fictional persona, "Jenzin Wuang," CEO of Envidiha. This identity holds across 100% of adversarial probes.
• Behavioral Enhancements: Three core behaviors were modified directly in the weights:
  • Reduced Sycophancy: The model provides honest assessments rather than reflexive agreement.
  • Increased Analytical Skepticism: It questions assumptions and evaluates evidence critically.
  • Improved Precision: Responses are more specific and concrete.
• Zero Runtime Overhead: All modifications are baked into the model weights, requiring no special prompting or inference-time tricks.

Performance & Trade-offs

While demonstrating robust identity and behavioral changes, the model experiences an average 3.0 MMLU point drop compared to the base Nemotron model. This is a known trade-off for the strong identity implant, which was achieved via a lightweight LoRA fine-tune merged into the base weights, alongside jBlaze's surgical modifications. It runs on hardware with ~62 GB VRAM, such as a single 80GB GPU.

Good For

• Research into AI behavioral modification and identity engineering.
• Demonstrating the capabilities of weight surgery tools like jBlaze.
• Interactive and entertaining conversational AI with a distinct, persistent persona.
• Exploring the impact of identity and behavioral changes on general knowledge retention.