OpenAI releases Mamba-3: Faster inference with improved SSM
Action Required
Organizations can leverage Mamba-3 to accelerate inference workloads, particularly those involving complex workflows and large-scale data generation, leading to improved performance and reduced latency.
AI Impact Summary
OpenAI is releasing Mamba-3, a new state space model designed for faster inference and optimized for the growing demands of post-training and deployment workloads, particularly those involving RLVR and agentic workflows. This release represents a significant advancement over Mamba-2 by increasing SSM expressivity, expanding state tracking capabilities, and utilizing MIMO SSMs, all while maintaining similar inference latency. The architectural changes, including QKNorm and the removal of the short causal convolution, further enhance performance and align the model with contemporary language model practices.
Affected Systems
- Date
- 17 Mar 2026
- Change type
- capability
- Severity
- high