Release 12Release 12 “Thebe” is here: GPU support, much better image controls, and a big dashboard update

Read the release notes

Why do I need Unikraft? I'll just grab Firecracker

Firecracker is a superb virtual machine monitor, but it is only one component of a full platform. Here is exactly what you get when you run Unikraft instead of building it yourself.

Felipe Huici
Felipe Huici
Co-Founder & CEO

Firecracker is superb! In fact, we run a hardened fork of it in production. But Firecracker is a virtual machine monitor: software that creates and runs virtual machines. A platform is everything wrapped around it like the guest, the snapshots, the orchestration, the density, and the team that keeps it patched. Here is exactly what that difference looks like.

  • Under 10 ms cold start for any workload
  • 100K+ microVMs per server
  • 10–100× better unit economics

A VMM does not make an entire platform

Grab Firecracker and you get the isolation boundary. Everything above it (click a layer on the right to see what it does) is what you would otherwise design, build, operate, and patch yourself.

If you grab Firecracker 1 component
Deploy · BYOC · on-prem · support
DIY
Kubernetes — Kraftlet virtual node
DIY
Orchestration — autoscale · live migration
DIY
Snapshot — fork · template · checkpoint
DIY
Persistent shared storage
DIY
Guest OS — minimal Linux
DIY
Firecracker VMM
you get this
KVM (Linux)
Hardware
The Unikraft platform full stack
Firecracker VMM (hardened fork)
shared
KVM (Linux)
Hardware

Deploy · BYOC · on-prem · support

Run the whole platform on bare metal or cloud VMs — on-prem, bring-your-own-cloud, or dedicated hosts. You own the full stack. SOC 2 and HIPAA compliant, with engineers on call.

Grab Firecracker and you get the VMM. The Unikraft platform provides everything above it — click a layer to see what it adds.

The Unikraft platform provides the full stack:

  • Deploy (BYOC, hosted, or on-prem) and get support: Run the whole platform on bare metal or cloud VMs, deploy on-prem, bring-your-own-cloud, or on dedicated hosts. You own the full stack. SOC 2 and HIPAA compliant, with engineers on call.
  • Kubernetes via Kraftlet virtual node: Drop Unikraft into an existing EKS/GKE/AKS cluster as a node. Schedule ordinary pods, but have them run as strongly-isolated microVMs that scale to zero in 10 ms.
  • Orchestration, autoscale, and live migration: Scale to zero and statefully resume in under 10 ms. Spin up a fleet of new hosts in under 20 s to absorb spikes. Live-migrate running VMs across servers for HA and maintenance.
  • Snapshot, fork, template, and checkpoint: Branch a live VM and its children in ~10 ms. Spawn thousands of instances from a template. Take named, restorable checkpoints, rewind a running VM in place, and launch new VMs from any point in history — all in milliseconds. Upstream Firecracker has none of this at this granularity.
  • Persistent shared storage: Cluster-wide volumes shared seamlessly by VMs running on different servers.
  • Guest OS (minimal Linux): A Dockerfile becomes your own trimmed-down Linux kernel. The guest boots in a few ms. A stock Linux guest on raw Firecracker needs ~330 ms+ just to reach userspace.
  • Hardened Firecracker fork: The isolation boundary: a fast, minimal device model with VM-grade security. This one layer is exactly where “just grab Firecracker” stops — and even here we run a hardened fork.

What’s actually included

Capability Raw Firecracker (DIY) Unikraft
Speed Cold start to ready
Boot a full guest OS, ~330 ms+
<10 ms, any workload
Speed Scale to zero + resume
Not included
Stateful resume <10 ms
Func Fork a live VM
Not included
~10 ms, children too
Func Templates — spawn N
Build it yourself
1,000s of templates
Func Checkpoint & restore
Basic pause/resume only
Named history + in-place rewind
Func Build from Dockerfile
Bring your own image
Dockerfile → microVM
Func Persistent shared storage
Not included
Cluster-shared volumes
Density Density per server
Your engineering
100K+ microVMs
Density Unit economics
Your engineering
10–100× better
Reliab Live migration
Not included
Across hosts, live
Reliab Autoscale hosts
Not included
New hosts <20 s
Reliab Kubernetes
Not included
Kraftlet virtual node
Reliab CVE tracking & patching
You own it
Maintained hardened fork
Reliab Support & compliance
Community OSS
Engineers + SOC 2 & HIPAA
Func Network shield / secret injection
Not included
Isolated shield microVM
Func Extend a running VM (plugins)
Build it yourself
Authenticated plugin API
"DIY" = possible with raw Firecracker, but you design, build, operate, and patch it yourself.

Functionality, speed, density, reliability

Speed — under 10 ms cold start

Instant boot for any workload and instant scale-to-zero.

  • Cold-start any workload in under 10 ms — and scale to zero, then statefully resume, in under 10 ms.
  • The microVM guest boots in a few ms on top of the VMM. A stock Linux guest on raw Firecracker needs ~330 ms+ just to reach userspace.

“We took Go codebases from ~4 s startup with Docker on GCP to ~30 ms on Unikraft.”

— Matthew Anderson, ex-Equinix Metal

Functionality — 10 ms to fork a live VM

A full snapshot suite Firecracker doesn’t have.

  • Fork a running VM (and its children) in ~10 ms - then spawn thousands of instances from a template.
  • Branch live workloads, take named checkpoints, rewind a VM in place, or launch new VMs from any point, all in millisecond timescales.
  • Turn any Dockerfile into your own minimal Linux microVM without rewrites, no wasm-specific changes.
  • Persistent shared storage, a plugin API, and custom filesystems across the cluster.

Density & economics — 100K+ microVMs per server

Density raw Firecracker alone won’t give you.

  • Run 100,000+ strongly-isolated microVMs on a single standard server.
  • 10–100× better unit economics.

“We run over 100,000 strongly isolated PostgreSQL instances on a single machine.”

— Sören Schmidt, CEO, Prisma

Reliability — under 20s to add a full host

  • Live-migrate running VMs across hosts for HA, maintenance, re-balancing, and spot instances.
  • Autoscale a fleet of new hosts in under 20 s for spikes — and scale back to zero when idle.
  • Run as a node in your existing k8s cluster (EKS/GKE/AKS) via Kraftlet — pods run as strongly-isolated microVMs.
  • SOC 2 Type II and HIPAA compliant. Deploy on-prem, BYOC, or dedicated — you own the full stack, with engineers on call.

Our Firecracker isn’t the one you’d grab.

If you grab the open-source VMM you will inherit the roadmap and the CVE queue. We maintain a large fork of Firecracker, and it ships on a monthly cadence. Release 11 “Metis” of our platform, just by itself, added branching & checkpointing, custom filesystems, a plugin API, and a network shield.

Snapshotting & state — Branch a live microVM (memory + volumes) with ~ms of downtime, or take named, restorable checkpoints. Custom filesystems give agents state that follows the work across the cluster.

Networking & the shield — VPC-style multi-interface networking, plus a transparent network shield: a separate scale-to-zero microVM that filters traffic and injects secrets — so untrusted workloads never hold a secret or get unfiltered egress.

The plugin API — Transparently extend a running microVM with your own code over an authenticated channel — in-VM command execution and filesystem services, no unauthenticated service exposed. It powers the native sandbox.

Platform & operations — A rebuilt dashboard with real-time metrics and logs, automatic disk reclamation, richer health checks for drain/upgrade workflows, and always-current online enterprise docs.

Also in the platform — more you’d otherwise build: GPU support, ARM, and dynamic vertical scaling (CPU & memory on the fly).

Don’t rebuild the platform. Deploy on it.

VM-grade isolation, 10 ms cold starts, and 100K+ instances per server — on-prem, BYOC, or dedicated. No setup, no commitment.