Mamba-3: A Linear-Complexity Alternative to Transformers

On March 17, 2026, a research team from Carnegie Mellon University, Princeton University, Cartesia AI and Together AI published a paper on Mamba-3, a new state space model (SSM) architecture designed primarily for inference efficiency rather than training efficiency, as its predecessor Mamba-2 was. The headline result: the Mamba-3 SISO variant outperforms Mamba-2, the Gated DeltaNet architecture and even the Llama-3.2-1B transformer on combined prefill and decode latency at every sequence length tested, at a model scale of 1.5 billion parameters.

Key facts about the release

• Publication date: March 17, 2026
• Authors: Aakash Lahoti and Kevin Y. Li (Carnegie Mellon University), Berlin Chen and Caitlin Wang (Princeton University), Aviv Bick and J. Zico Kolter (Carnegie Mellon University), Tri Dao (Princeton University, Together AI), Albert Gu (Carnegie Mellon University, Cartesia AI)
• Test scale: models of about 1.5 billion parameters
• Key architectural changes: a more expressive recurrence formula, complex-valued state tracking and a MIMO (multi-input, multi-output) variant
• Open source: compute kernels in Triton, TileLang and CuTe DSL are publicly available

Why does Mamba-2 no longer meet the industry's needs?

Since Mamba-2 came out in mid-2024, most SSM-based architectures have moved to it from Mamba-1. Its developers bet that training speed was still the main bottleneck, and they simplified the core SSM mechanism to get a 2–8x training speedup over its predecessor. The bet paid off and Mamba-2 was widely adopted across the industry.

But the large language model landscape has shifted noticeably since then. Pretraining still matters, but far more attention now goes to post-training and deployment, both of which put heavy demands on inference. Scaling post-training methods, especially reinforcement learning with verifiable rewards (RLVR) for coding and math, requires generating huge numbers of rollouts, while agentic workflows such as Codex, Claude Code or OpenClaw have sharply increased demand for inference.

Even so, many linear architectures, Mamba-2 included, were originally designed with training as the priority. To speed up pretraining, the core SSM mechanism was progressively simplified. For example, the diagonal state transition was reduced to a scalar multiplied by an identity matrix, which cut the model's expressiveness in exchange for faster training.

What does

Mamba-3 change, and why build an inference-optimized architecture?

Mamba-3 reverses this logic. The team restored a more expressive recurrence formula and complex-valued state tracking, mechanisms that Mamba-2 had simplified for training speed, and added a MIMO variant that improves accuracy without slowing decoding. The result is an architecture with lower inference latency than Mamba-2, Gated DeltaNet and even the classic Llama-3.2-1B transformer, with the advantage holding at every sequence length tested at 1.5 billion parameters. Because SSM architectures process a sequence with linear computational complexity, rather than the quadratic complexity of transformers, Mamba-3's gain is most visible on long contexts and under heavy generation load, which is exactly where the industry's demand for fast, cheap inference is concentrated today.

What this means for the industry

The open release of compute kernels in Triton, TileLang and CuTe DSL lowers the barrier for teams that want to bring Mamba-3 into their own inference infrastructure. The collaboration between academic labs (Carnegie Mellon University, Princeton University) and industry players (Together AI, Cartesia AI) reflects a broader trend: cutting-edge architecture research increasingly ships with ready-made infrastructure for practical deployment, rather than staying a theoretical result on paper. For the market, it is a signal that linear-complexity alternatives to transformers are moving from lab experiments to real candidates for the base architecture of next-generation inference-oriented products, from agentic assistants to large-scale post-training services.