Microsoft AI released MAI-Cyber-1-Flash, a 137B total parameter model with 5B active parameters built as a sparse MoE fine-tune of MAI-Code-1-Flash, designed to run within Microsoft’s MDASH multi-agent system for cyber defense (MarkTechPost). The model achieves 95.95% on CyberGym and includes a 256k context window (The Decoder). According to Microsoft, the system reduces costs by approximately 50 percent compared to pure frontier models by routing only complex cases to GPT-5.4 (The Decoder).
Moonshot AI released Kimi K3’s model weights alongside open-source infrastructure components (The Decoder). The company and kvcache-ai open-sourced AgentENV under MIT license, a distributed system that powers agentic reinforcement learning training by running agent sandboxes as Firecracker microVMs with millisecond snapshot, resume, and 16-way fork capabilities behind an E2B-compatible API (MarkTechPost). Perplexity released pplx, a single-binary CLI tool exposing its Search API through terminal commands for coding agents, shipping as a checksum-verified binary for macOS arm64 and Linux with an Agent Skills module (MarkTechPost).