UncannyEcho/Aura-v1.1-BF16
Aura v1.1 BF16 is a 7.9 billion parameter language model developed by UncannyEcho, based on Google's unquantized Gemma 4 E4B QAT checkpoint. This half-precision model features a refined personality adapter and a 32768 token context length, making it suitable for high-fidelity Transformers-based inference, evaluation, and research. It excels in natural conversation, creative writing, reasoning, and multimodal tasks, with a focus on on-device deployment.
Loading preview...
Aura v1.1 BF16 Overview
Aura v1.1 BF16 is a 7.9 billion parameter language model designed for high-fidelity Transformers inference, evaluation, and research. It is a half-precision (BFloat16) variant, built upon Google's unquantized Gemma 4 E4B QAT checkpoint and incorporates a canonical combined Aura adapter. This release features significant improvements to its personality adapter, merged with the same ablation adapter as its predecessor.
Key Capabilities
- High Fidelity: Optimized for precise and nuanced outputs.
- Foundation: Derived from a robust, unquantized Gemma 4 E4B QAT checkpoint.
- Versatile Deployment: Intended for local GPU workstations and servers, supporting Transformers-based runtime.
- Broad Task Support: Capable of natural conversation, creative writing, reasoning, and multimodal tasks.
- Enhanced Personality: Features an improved personality adapter for more engaging interactions.
Good For
- Transformers-based Inference: Ideal for applications requiring high-quality language generation.
- Evaluation and Research: Suitable for academic and developmental purposes in LLM studies.
- On-Device Deployment: Designed with considerations for efficient operation on local hardware.
- Companion/Friend Applications: Can be utilized in scenarios where a conversational AI with personality is desired.
- Multimodal Applications: Supports tasks that integrate various data types beyond text.