Aura-2-Thinking

Aura-2-Thinking is an 11B parameter reasoning model based on google/flan-t5-xxl, fine-tuned to output step-by-step thinking traces inside <think> tags before providing the final answer.

Model Details

• Architecture: Flan-T5-XXL (Encoder-Decoder)
• Parameters: ~11.15 Billion
• Format: Standalone un-sharded model.safetensors (~10.62 GB)
• Precision: F8_E4M3 (torch.float8_e4m3fn)
• Primary Use: Chain-of-thought reasoning, multi-step problem solving, math/logic tasks