Mamba-3: transformer alternative with linear complexity

On March 17, 2026, a research group from Carnegie Mellon University, Princeton University, Cartesia AI, and Together AI published a paper on a new Mamba-3 architecture — a state space model (SSM) designed primarily for inference efficiency, not training, unlike its predecessor Mamba-2. The main result: the Mamba-3 SISO variant outperforms Mamba-2, the Gated DeltaNet architecture, and even the Llama-3.2-1B transformer in total latency at prefill and decode stages across all sequence lengths 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)
• Scale of tests — models at 1.5 billion parameter level
• Key architectural innovations — more expressive recurrence formula, complex-valued state tracking, and MIMO variant (multi-input, multi-output)
• Open source — computational kernels in Triton, TileLang, and CuTe DSL published in open access

Why Mamba-2 no longer meets industry needs?

Since the release of Mamba-2 in mid-2024, most SSM-based architectures have switched to it from Mamba-1: developers bet that the main bottleneck remained training speed, and simplified the basic SSM mechanism for 2–8x training acceleration compared to its predecessor — this bet paid off and ensured Mamba-2 wide adoption in the industry. But since then, the landscape of large language models has shifted noticeably: although pretraining remains important, far more attention has been paid to post-training and deployment — both processes that are extremely demanding on inference. Scaling post-training methods, especially reinforcement learning on verifiable rewards (RLVR) for coding and math tasks, requires generating huge numbers of rollouts, and agentic workflows — such as Codex, Claude Code, or OpenClaw — have sharply increased demand for inference.

Despite this, many linear architectures, including Mamba-2, were originally developed with a priority on training: to speed up pretraining, the basic SSM mechanism was progressively simplified — for example, the state transition was reduced to a scalar multiplied by an identity matrix, reducing model expressiveness for training speed.

What

Mamba-3 changes and why does an architecture optimized for inference matter?

Mamba-3 reverses this logic: the team restored to the model a more expressive recurrence formula and complex-valued state tracking — mechanisms that Mamba-2 simplified for training speed — and also added a MIMO variant that increases accuracy without slowing decoding. The result — an architecture with inference latency lower than Mamba-2, Gated DeltaNet, and even the classical Llama-3.2-1B transformer, with the advantage holding across all tested sequence lengths at a scale of 1.

5 billion parameters. Because SSM architectures process sequences with linear rather than quadratic computational complexity as in transformers, Mamba-3's advantage is particularly noticeable on long contexts and under high load at the generation stage — exactly where today's main industry demand for fast and cheap inference is concentrated.

What this means for the industry

The open publication of computational kernels in Triton, TileLang, and CuTe DSL lowers the barrier to entry for teams wanting to implement Mamba-3 in their own inference infrastructure, and collaboration between academic laboratories (Carnegie Mellon University, Princeton University) and industrial players (Together AI, Cartesia AI) reflects a broader trend: cutting-edge architectural research is increasingly accompanied right away by ready infrastructure for practical deployment, rather than remaining only a theoretical result on paper. For the market, this signals that alternatives to transformers with linear computational complexity are transitioning from the category of laboratory experiments to the category of real candidates for the role of base architecture for inference-oriented products of the next generation — from agentic assistants to large-scale post-training services.