smol machines 2026: shipping a whole Linux VM that boots in under 200ms

smol machines builds smolvm, an open-source CLI and SDK that boots Linux microVMs with their own guest kernel on macOS, Linux, or Windows with sub-second cold starts, packs stateful machines into single-file .smolmachine artifacts, and runs the same artifact in a hosted cloud.

Developers will ship a whole hardware-isolated VM instead of a container if it boots in under 200ms and packs into one portable file, not just an image.

BinBin He (HN: binsquare), founder of the YC Spring 2026 startup, started from the 'works on my machine' problem: containers isolate weakly because they share the host kernel, while real VMs take seconds to boot. The project aimed to make hardware virtualization as cheap and portable as a container, and the 2026-04-17 Show HN introduced sub-second cold starts plus .smolmachine artifacts.

The Show HN drew 499 points and 152 comments, with discussion centered on sandboxing untrusted code, AI agents, and running Docker or Kubernetes inside microVMs. The company joined Y Combinator's Spring 2026 batch, and by mid-2026 shipped a hosted smol cloud so the same SDK and .smolmachine artifact run locally, in the cloud, or self-hosted, along with GPU access (Vulkan) and CUDA API remoting.

Active as of 2026-09-02: open-source runtime on GitHub plus a hosted cloud, positioned for AI coding agents and untrusted-code sandboxing rather than as a drop-in Docker replacement.

Containers share the host kernel, so their isolation boundary is weaker than the machine boundary developers already trust.

Classic VMs were correct but slow to boot, which is why developers tolerated weaker container isolation.

Booting OCI images with no daemon and packing whole machines into one file made the strong option convenient.

AI agents and untrusted-code execution created a moment where hardware isolation was newly in demand.

A portable full VM also solved environment drift, not just security, giving developers two reasons to switch.

Making the strongest isolation cheap - a sub-200ms full VM - turns 'works on my machine' and 'untrusted code' into product wedges instead of hard constraints.

As of 2026-09-02 smol machines is an active YC Spring 2026 company: smolvm is Apache-2.0 on GitHub, with a hosted smol cloud and self-hosted deployment of the same .smolmachine artifacts. Engineering posts show workloads containers struggle with - a real Kubernetes cluster and GPU-accelerated rendering inside one machine - and the project targets AI coding agents and automated sandboxing. Limitations are documented openly (Windows lacks GPU support and pack creation), and the security claim is bounded: hypervisor isolation per workload, not a hardened multi-tenant control plane.