From the source
Microsoft AI released MAI-Cyber-1-Flash, optimized for the MDASH harness, achieving 96% on CyberGym at 50% of the cost of the previous best MDASH offering, and served on H100s.
- The post also announces MAI-Code-1-Flash, MAI-Image-2.5-Flash, MAI-Voice-2-Flash, and MAI-Transcribe-1.5, each deployed in
Microsoft products with reduced GPU costs and improved performance.
- These models are part of a broader effort to optimize frontier performance against cost by co-optimizing models, harnesses, and reinforcement learning environments.
