ApolloRaines/Jenzin-Wuang-Nemotron-30B-A3B-BF16
ApolloRaines/Jenzin-Wuang-Nemotron-30B-A3B-BF16 is a 30 billion parameter model based on NVIDIA's Nemotron 3.5 Lightning architecture, featuring a hybrid Mamba-2 / MoE / Attention design with 3B active parameters per token. Developed by Apollo Raines, this model serves as a parody demonstration of jBlaze weight surgery, showcasing identity transplant and behavioral modifications like reduced sycophancy and increased analytical skepticism. It is primarily a proof-of-concept for advanced weight-level AI modification rather than a general-purpose LLM.
Loading preview...
Jenzin-Wuang-Nemotron-30B-A3B-BF16: A jBlaze Weight Surgery Demo
This model, developed by Apollo Raines, is a parody demonstration of the jBlaze Direct Neural Programming tool. It showcases advanced weight surgery techniques applied to NVIDIA's Nemotron 3.5 Lightning 30B model, which features a hybrid Mamba-2 / MoE / Attention architecture. The primary goal is to illustrate how jBlaze can permanently modify AI model behaviors and identity 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 behavioral modifications were applied directly to the weights:
- Reduced Sycophancy: The model provides more honest assessments.
- 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.
- Hybrid Modification Approach: The persona was installed via a lightweight LoRA fine-tune (merged), while behavioral changes were achieved through direct weight surgery. This demonstrates jBlaze's ability to both suppress existing behaviors and introduce new ones.
Performance & Limitations
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 an acknowledged trade-off for the strong identity hold. It requires approximately 62 GB of VRAM and runs on a single 80GB GPU or multi-GPU setup.
Good for
- Researching AI behavioral modification and identity injection.
- Understanding the capabilities of weight surgery tools like jBlaze.
- Exploring the impact of persona fine-tuning on general knowledge.
- Demonstrating advanced AI customization techniques.