Setsuna · Currently in research preview
16 GiB microVM88 milliseconds
True-cold Firecracker boots. No snapshots or warm pools.
Setsuna Benchmarks
AI agents increasingly need isolated CPU for coding, evals, browser execution, and untrusted workloads. Existing isolation techniques still battle with the decision between security and startup latency. Setsuna removes that tradeoff: true-cold Firecracker microVMs that become usable fast enough to serve as an on-demand primitive.
Our published results are based on 500 and 1000 sequentially run benchmarks per machine-specific host and guest configuration that we continuously optimize. Benchmark proof is publicly available for download. We encourage you to reach out for access to the research preview.
Machine | Configurations | |||
|---|---|---|---|---|
| Azure Standard_F8amds_v7 | 6 | 82 ms | 75 ms | 145 ms |
| Azure Standard_E16ads_v7 | 4 | 85 ms | 77 ms | 112 ms |
| Azure Standard_D8ads_v7 | 4 | 87 ms | 80 ms | 201 ms |
| Azure Standard_E8ads_v7 | 1 | 96 ms | 83 ms | 113 ms |
| Azure Standard_F16alds_v7 | 1 | 117 ms | 104 ms | 134 ms |
| Azure Standard_E8ads_v6 | 4 | 145 ms | 121 ms | 276 ms |
| Azure Standard_F16ads_v7 | 1 | 145 ms | 131 ms | 159 ms |
| Azure Standard_D32ads_v7 | 1 | 293 ms | 271 ms | 375 ms |
Page 1 of 1
Setsuna VS Others
| Capability | Setsuna | E2B | Modal | Fly Machines | Vercel Sandbox |
|---|---|---|---|---|---|
| Cold-start contract | |||||
| True cold boot on every start | [2] | [1] | [3] | — | |
| No snapshot resume | [2] | [1] | [3] | [1] | |
| No warm pool | — | — | — | ||
| No preboot | [2] | — | — | — | |
| Documented performance | |||||
| Request → ready, p50 | 88 ms | ~717 ms[1] | ~2.4 s[1] | ~2+ s[3] | ~1.9 s[1] |
| Request → ready, p99 | 99 ms | — | — | — | — |
| Published sample size | 1,000 runs | — | — | — | — |
| Raw samples downloadable | |||||
| Checksummed proof bundle | |||||
| Isolation | |||||
| Isolation model | Firecracker microVM | Firecracker (snapshot resume)[2] | gVisor[1] | Firecracker microVM[3] | Firecracker microVM[1] |
| Measured configuration | |||||
| Guest resources disclosed | 2 vCPU · 16 GB | — | — | — | — |
| Host hardware disclosed | |||||
Bracketed numbers point to the sources below. Competitor create/start figures are measured values from LogRocket’s independent August 2026 benchmark or the vendor’s own documentation; lifecycle spans differ per platform and several are snapshot-backed rather than true-cold boots. A red cross means the capability is not offered, per the cited source. A dash means the cited sources do not document it either way.
Request to strict-ready: Host-local wall-clock time from the start of the sandbox create request until the workspace is strictly verified ready — not just kernel start.
- [1] LogRocket — “Comparing AI agent sandbox platforms: E2B, Modal, Daytona, and more” · August 2026
- [2] E2B — infrastructure architecture (docs/ARCHITECTURE.md) · retrieved August 2026
- [3] Fly.io — Machine suspend and resume · retrieved August 2026
Ready. SetSuna.
Setsuna (刹那) is a Japanese word born from the ancient Buddhist text Abhidharmakośa-śāstra, where it measured the physical universe as a concrete unit of time. Its literal meaning? “A split second.” But because a setsuna is so mind-blowingly brief, philosophers used it as the ultimate metaphor for change: a reminder that nothing stays the same, and everything is constantly being reborn in every single instant.
We applied that same philosophy of quick, continuous rebirth to Firecracker microVMs. Every time compute is requested, a genuinely new machine is born from nothing, executes its workload, and vanishes.
- No past lives
- Zero stale state. Every environment runs on a new VMM process, a new KVM guest, and a fresh kernel boot. There are no snapshots to restore, no warm pools to draw from, and no fleets of idle memory padding the bill.
- Sub-100 ms rebirth
- The accepted penalty for true hardware isolation has always been startup latency. Setsuna breaks that bottleneck, booting a full 16 GiB machine from nothing to verified readiness with a p50 of 88 milliseconds.
- Clean-slate execution
- A pristine, CPU-enforced boundary per task stops being an architectural luxury. It becomes the baseline for secure, disposable compute, without the historical latency tax of a cold boot.
Research preview
Request access to Setsuna microVM.
Our benchmark results, methodology, and evidence are published online. Setsuna is currently limited to selected evaluations and not generally available.
Setsuna is currently available for selected enterprise, infrastructure, and research evaluations.