QuantaSparkLabs/NYXIS-Pro
NYXIS-Pro by QuantaSparkLabs is a 1.1 billion parameter language model built upon the NYXIS1.1B base, enhanced with a curated 5,000+ chunk AEGIS knowledge base. It features a dual-stage retrieval system for factual grounding, ensuring zero hallucination by relying on verified data. The model also incorporates an Emotion Engine to detect user sentiment and adjust its response tone. NYXIS-Pro is optimized for factual question answering across coding, math, science, and general knowledge, prioritizing accuracy and honesty.
Loading preview...
NYXIS-Pro: The AEGIS Knowledge Shield
NYXIS-Pro, developed by QuantaSparkLabs, is a 1.1 billion parameter model designed for highly accurate and grounded responses. It integrates the NYXIS1.1B base model with a specialized AEGIS Knowledge Base containing over 5,000 curated chunks from sources like Wikipedia, Trivia QA, and ArXiv, covering diverse domains including coding, math, science, and history.
Key Capabilities & Differentiators
- Zero Hallucination: Employs a Dual-Stage Retriever (bge-small for recall, ms-marco-MiniLM for reranking) to ground all factual answers in verified data, explicitly stating when information is unavailable.
- 4-Tier Knowledge Cascade: Prioritizes internal AEGIS knowledge, then external search (user-provided API), followed by its model brain, and finally an honest fallback.
- Emotion Engine: Detects user emotional states on a 1-10 scale and dynamically adjusts its response style to match the user's tone, offering empathetic, balanced, or energetic interactions.
- Optimized for Factual Accuracy: Its architecture is specifically engineered to prevent fabrication, making it reliable for information retrieval.
Use Cases & Limitations
NYXIS-Pro is ideal for applications requiring high factual accuracy and grounded responses, particularly in educational, technical support, or information-seeking contexts where hallucination is unacceptable. It requires approximately 2.1 GB VRAM (4-bit) and is English-only. Its knowledge is limited to the ~5,400 chunks in its AEGIS base, and its emotion detection is keyword-based rather than deep sentiment analysis.