Open source Mamba 3 arrives to surpass Transformer architecture with nearly 4% improved language modeling, reduced latency

Got story updates? Submit your updates here. ›

The generative AI era began with the launch of OpenAI's ChatGPT in late 2022, but the underlying Transformer neural network architecture dates back to Google's 2017 paper. While Transformers deliver unparalleled model quality, they are computationally expensive. Researchers have now released Mamba-3, an open-source language model that improves on Transformer architecture with nearly 4% better language modeling and reduced latency.

Why it matters

Mamba-3 represents a strategic shift in the total cost of ownership for AI deployments, doubling inference throughput for the same hardware footprint. As organizations move toward parallel, agentic workflows, the demand for low-latency generation increases, and Mamba-3 is designed to prevent GPU hardware from sitting idle during these tasks.

The details

Mamba-3 is a type of State Space Model (SSM) that maintains a compact, ever-changing internal state to process information quickly with lower memory requirements. It achieves comparable perplexity to its predecessor, Mamba-2, while using only half the state size. Mamba-3 introduces three key technological leaps: Exponential-Trapezoidal Discretization, Complex-Valued SSMs and the "RoPE Trick", and Multi-Input, Multi-Output (MIMO) to boost arithmetic intensity.

• Mamba-3 was released in March 2026.

What they’re saying