Quick Facts
- Mamba-3 achieves 57.6% average accuracy at 1.5B parameters, a 2.2-percentage-point improvement over Transformers
- The new architecture doubles inference throughput while using half the memory of previous versions
- Model is available under Apache-2.0 license with code published on GitHub for immediate deployment
Researchers from Carnegie Mellon University and Princeton University released Mamba-3 in March 2026, an open-source state space model that outperforms the industry-standard Transformer architecture while requiring significantly less computational resources.
The new model achieved a 57.6% average accuracy across benchmarks at the 1.5-billion-parameter scale. This represents a 2.2-percentage-point improvement over Transformers, translating to nearly a 4% relative increase in language modeling capability.
Mamba-3 delivers the same performance as its predecessor while using only half the internal state size. This efficiency gain doubles inference throughput for the same hardware footprint, addressing a critical cost concern for AI deployments.
The architecture beats Mamba-2, Gated DeltaNet, and Llama-3.2-1B on prefill and decode latency across all sequence lengths. The SISO variant achieves the fastest performance in this category, while the MIMO variant matches Mamba-2’s speed with substantially stronger accuracy.
Albert Gu of Carnegie Mellon and Tri Dao of Princeton led the development team. The project received support from Schmidt Sciences AI2050 fellowship, Google ML and Systems Junior Faculty Awards, and commercial partners Together AI and Cartesia AI.
The research team implemented three core improvements: enhanced SSM recurrence with complex-valued exponential-trapezoidal formulas, complex-valued state tracking, and updated normalization methods that stabilize training.
Together AI demonstrated commercial success with previous Mamba architectures, achieving industry-leading latency under 200 milliseconds and 2x faster performance at half the cost of competing providers. The company recently raised funding at a $3.3 billion valuation.
Mamba-3 addresses the computational limitations of Transformers, which require quadratic compute and linear memory demands that make large-scale inference expensive. The new architecture prevents GPU hardware from sitting idle during parallel, agentic workflows.
The model targets organizations building long-context applications and real-time reasoning agents. As the AI industry shifts focus from pretraining to deployment and post-training optimization, inference efficiency becomes increasingly critical for commercial viability.
The complete model code and optimized kernels are available on GitHub under the Apache-2.0 license, built using Triton, TileLang, and CuTe DSL for maximum hardware performance.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.