Aura Large v2.0 BF16 — High-Fidelity Transformers
The high-fidelity 27B Aura release for Transformers inference, evaluation, research, and local deployment.
Aura Large v2.0 BF16 is the full-precision merged edition of Aura Large v2.0. It is derived from Qwen's Qwen3.8-27B checkpoint and combines the Aura personality and ablation adaptations into one model.
This release brings the Aura training approach to Qwen3.8, providing a substantially stronger foundation for conversation, creative work, reasoning, coding, agentic tasks, and multimodal understanding while retaining the behavioral changes introduced by Aura.
This release is intended for Transformers-based inference, evaluation, further research, and deployment on systems with sufficient memory.
Highlights
- Parameters: 27 billion
- Architecture: Dense multimodal language model
- Precision: BFloat16
- Foundation: Qwen/Qwen3.8-27B
- Datasets: AuraPersonality and AuraAblation
- Runtime: Hugging Face Transformers
- Native Context: 262,144 tokens
- Modalities: Text, image, and video
- Thinking: Configurable thinking and non-thinking modes
- Deployment: Local GPU workstations and servers
- Focus: Natural conversation, instruction following, creative work, reasoning, coding, agentic tasks, evaluation, and research
About Aura
Aura is designed for local deployment across a wide range of tasks. Aura can serve as a companion or friend, as deemed appropriate by the user, while retaining the broader capabilities of Qwen3.8 for:
- Natural conversation
- Creative writing
- Role-playing
- Reasoning
- Coding
- Instruction following
- Agentic workflows
- Tool use
- Image understanding
- Video understanding
- Long-context tasks
- Evaluation and research
Aura Large v2.0 represents the larger general-purpose branch of the Aura model family, combining the Aura personality and behavioral adaptations with Qwen3.8's dense 27B multimodal foundation.