Serverless Firecracker VMs that snapshot, clone, and migrate

What it does

Tensorlake is a cloud platform built around two ideas: stateful Firecracker MicroVM sandboxes for isolated agent execution, and a serverless orchestration runtime for chaining those sandboxes into background workflows. You can spin up a sandbox, run code inside it, snapshot its memory and filesystem, clone it to another machine, or let it auto-suspend when idle. The orchestration layer lets you deploy Python functions that each execute inside their own sandbox and fan out across a distributed cluster.

The interesting bit

Most serverless offerings treat state as baggage; Tensorlake treats long-lived VM state as a first-class primitive. Sandboxes can live-migrate during infrastructure updates, and the README claims a single project can host five million of them. The block-based storage is benchmarked with SQLite on a 2 vCPU / 4 GB RAM setup, finishing in 2.45 seconds against competitors ranging from 3.00 to 5.51 seconds.

Key highlights

• Sandboxes boot in under a second via a custom scheduler called Lattice.
• Snapshot and clone: checkpoint a running VM’s memory and disk, then instantiate copies elsewhere.
• Auto-suspend and resume: VMs hibernate when idle and wake in under a second with state intact.
• Live migration: sandboxes move between physical hosts during updates with only a brief pause.
• Orchestration supports fan-out and HTTP or SSE invocation for agent workflows.

Caveats

• Requires a Tensorlake Cloud account and API key; the README shows no self-hosted or open-source backend.
• The SQLite benchmark is a single data point on a specific 2 vCPU / 4 GB RAM configuration.
• The orchestration SDK and CLI are tightly coupled to the managed cloud service.

Verdict

Worth evaluating if you are building agents that need fast, isolated code execution with VM-level statefulness. Skip it if you require an on-premise or fully open-source sandbox runtime.