Moonshot open-sources the microVM engine that trained its 2.8T-param model

The infrastructure that trained one of the largest mixture-of-experts models on the planet is now on GitHub under an MIT license. Moonshot AI’s Kimi team and kvcache-ai released AgentENV on Friday, a distributed platform that spins up isolated Linux sandboxes for agentic reinforcement learning at scale. This is the same system that powered training for Kimi K3, Moonshot’s 2.8-trillion-parameter model.

Agentic RL is a different beast from text sampling. Every training rollout needs a real computer—filesystem, network stack, live processes and all. Containers are fast but share a kernel, which is a nonstarter when your model is generating and executing arbitrary code. Full VMs solve the isolation problem but boot like molasses and hog memory while idle. AgentENV splits the difference with Firecracker microVMs, then layers on snapshot, pause, resume, and fork primitives that make the whole thing fast enough for training workloads.

The numbers are what sell it. Snapshot-backed environments boot or resume in under 50 milliseconds. Pause takes under 100 milliseconds. Incremental snapshots capture in under 100 milliseconds even when the disk is getting hammered. And fork—the feature most directly relevant to RL—lets a running sandbox clone into up to 16 independent children on the same node, with the source pausing only briefly during capture. Expensive setup runs once, then branches into parallel rollouts.

There’s a smart adoption play here too. AgentENV ships with an E2B-compatible HTTP API, so teams already building on E2B’s hosted sandboxes can point their existing Python or TypeScript SDK at a self-hosted AgentENV server and keep moving. The code runs on Linux with kernel 6.8+ and /dev/kvm access, with deployment options ranging from a single systemd service to a full Kubernetes stack. The docs are refreshingly honest about requirements: at least 1 Gbps for shared storage, with 10 Gbps strongly recommended if you’re doing this for real.