Moonshot Open-Sources AgentENV, the Firecracker microVM Fleet That Trained Kimi K3's Agent RL

kvcache-ai / Moonshot AI (GitHub)·high signalXBlueskyLinkedInCopy link

Alongside the K3 weights, Moonshot's Kimi team and kvcache-ai released AgentENV (AENV) under an MIT license on July 27 — the distributed sandbox platform behind K3's agentic RL training. Each agent environment is a Firecracker microVM with its own Linux kernel, filesystem, and network namespace rather than a container, with overlaybd layered images sharing read-only base layers; the project reports sub-50ms boot/resume from snapshot, sub-100ms pause, and sub-100ms incremental snapshot capture, with a running sandbox forking into up to 16 children on one node. For builders, this is the concrete answer to 'containers share the host kernel and that isn't enough isolation for model-generated code' — a production-scale reference implementation you can actually read.