Aura v1 BF16 — High-Fidelity Transformers
The high-fidelity Aura release for Transformers inference, evaluation, and research.
Aura v1 BF16 is the half-precision edition of Aura, derived from Google’s unquantized Gemma 4 E4B QAT checkpoint and the same canonical combined Aura adapter used across the Aura v1 family.
This release is intended for Transformers-based inference, evaluation, further research, and deployment on GPUs with sufficient memory.
Highlights
- Precision: BFloat16
- Foundation: Unquantized Gemma 4 E4B QAT checkpoint
- Runtime: Hugging Face Transformers
- Deployment: Local GPU workstations and servers
- Focus: Fidelity, evaluation, and research
About Aura
Aura is designed for on-device deployment across multiple tasks. Aura can be a companion or friend, as deemed necessary by the user, while retaining the broader capabilities of Gemma 4 for:
- Natural conversation
- Creative writing
- Reasoning
- Multimodal tasks
- Evaluation and research