BAMBA-9B-V2

Summary from the TLDR AI newsletter

IBM, Princeton, CMU, and UIUC released Bamba v2, a Mamba2-based model that outperforms Llama 3.1 8B after training on 3 trillion tokens. Bamba v2 leverages Mamba2 architecture for 2-2.5x faster inference and excels in L1 and L2 benchmarks. The team plans further optimizations and invites community collaboration to enhance model performance.