New Mamba-2 Hybrid Architecture Challenges Transformer Dominance

Researchers at Carnegie Mellon University and Anthropic have published a paper introducing Mamba-2 Hybrid, a new architecture that combines state-space model (SSM) layers with selective attention layers to achieve transformer-level quality at dramatically lower inference costs.

The architecture uses Mamba-2 SSM layers for 80% of the model's depth, with strategically placed attention layers handling tasks that require precise long-range information retrieval.

This hybrid approach retains the transformer's ability to perform exact recall and complex reasoning while gaining the SSM's linear-time inference efficiency.

In benchmarks, a 7B parameter Mamba-2 Hybrid model matches the quality of a 13B parameter transformer while running 5x faster at inference and using 3x less memory.

The efficiency gains are most pronounced for long-context tasks, where the architecture's linear scaling replaces the transformer's quadratic attention computation.

Several AI companies have expressed interest in adopting the architecture. DeepSeek has already begun training a Mamba-2 Hybrid variant, and Meta's AI research team is evaluating it for future Llama iterations.

If the architecture proves reliable at frontier scale, it could significantly reduce the cost of running large language models, benefiting both cloud providers and local AI deployment through tools like Ollama.

The paper has been accepted at ICML 2026 and the code has been open-sourced on GitHub, enabling the research community to experiment with the architecture.

Anthropic's involvement in the research signals its interest in efficiency improvements alongside capability advances, potentially informing future Claude model architectures.