Ishowbackup/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-heretic
Ishowbackup/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-heretic is a 30 billion parameter decensored version of NVIDIA's Nemotron-3-Nano-30B-A3B-BF16, created using the Heretic v1.2.0 tool with Arbitrary-Rank Ablation. This model features a hybrid Mixture-of-Experts (MoE) architecture, combining Mamba-2 and Attention layers, and is designed for both reasoning and non-reasoning tasks with configurable reasoning traces. It supports English, German, Spanish, French, Italian, and Japanese, and excels in agentic and long-context tasks, demonstrating significantly reduced refusals compared to its original counterpart.
Loading preview...
Model Overview
This model, Ishowbackup/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16-heretic, is a decensored variant of NVIDIA's Nemotron-3-Nano-30B-A3B-BF16, processed with the Heretic v1.2.0 tool using Arbitrary-Rank Ablation (ARA). The original NVIDIA model is a 30 billion parameter large language model (LLM) developed by NVIDIA Corporation, featuring a unique hybrid Mixture-of-Experts (MoE) architecture that integrates Mamba-2 and Attention layers. It is designed for unified reasoning and non-reasoning tasks, capable of generating reasoning traces before providing a final response, which can be configured for higher accuracy on complex prompts.
Key Differentiators
- Decensored Version: Significantly reduces refusals from 99/100 in the original model to 6/100, offering broader utility.
- Hybrid MoE Architecture: Combines 23 Mamba-2 and MoE layers with 6 Attention layers, including 128 experts plus 1 shared expert per MoE layer, with 6 experts activated per token.
- Configurable Reasoning: Users can enable or disable intermediate reasoning traces, balancing between direct answers and higher accuracy for challenging tasks.
- Extensive Multilingual Support: Supports English, German, Spanish, French, Italian, and Japanese, with improved performance using Qwen.
- Long Context Window: Supports a maximum input and output size of 1 million tokens, though default Hugging Face configuration uses 256k tokens due to VRAM requirements.
Performance Highlights
- Reduced Refusals: Achieves 6/100 refusals compared to 99/100 in the original model.
- Strong Agentic Performance: Outperforms other models in benchmarks like AIME25 (with tools) at 99.2%, MiniF2F pass@1 at 50.0%, and SWE-Bench (OpenHands) at 38.8%.
- Long Context Capability: Achieves high scores on RULER-100 benchmarks, including 92.9% at 256k and 91.3% at 512k context lengths.
Ideal Use Cases
- AI Agent Systems: Well-suited for developing sophisticated AI agents that require robust reasoning capabilities.
- Chatbots and RAG Systems: Effective for building advanced conversational AI and retrieval-augmented generation applications.
- Instruction Following: Capable of handling typical instruction-following tasks across various domains.
- Commercial Applications: Ready for commercial deployment under the NVIDIA Nemotron Open Model License.