spark-vm vs Modal
Modal is serverless compute with a sandbox tier: scale-to-zero, GPUs inside the sandbox, memory snapshots — and, since October 2026, metered network egress. It's built for developers with workloads that burst. spark-vm is built for the agent that lives somewhere: a persistent computer for Muse, Grok Bot, Dots, and everything like them. Different customer, different job.
| spark-vm | Modal | |
|---|---|---|
| What it is | A persistent computer for an agent you own: files, jobs, desktop survive to tomorrow. | Serverless compute for developers (gVisor sandboxes) with GPUs, scale-to-zero, and memory snapshots. |
| Persistence model | Always-on by design. State is the default. | Snapshots + scale-to-zero. Memory snapshots exist; the unit still sleeps and wakes per workload. |
| Price shape | Self-host: free (MIT) + your hardware. Hosted: TBD, priced per box not per seat. | Free Starter · $250/mo Team. Sandbox tier ≈3× the standard compute rate (≈$0.0710/vCPU-hr equivalent). Egress billed since Oct 1, 2026: $0.04/GiB over allowance. |
| GPU | No GPU tier today. CPU-only, stated plainly. | GPUs inside the sandbox (T4 through B300 classes) — Modal's home turf. |
| Idle economics | The box is yours; idle costs whatever your hardware or flat box price costs. | Scale-to-zero. You don't pay for idle compute — but watch the new egress meter if your agent moves data. |
| Secrets handling | Credential proxy. The agent only ever sees hsurr: placeholders; swaps on allowlisted hosts, audited. |
Standard secrets management in the Modal environment. |
| Human approvals | Built in. Sensitive actions pause for a tap on your phone. | Yours to build. No per-action approval plane in the product. |
| Self-host option | Yes. MIT open source; your hardware, your tailnet, today. | No. Hosted platform only. |
When to pick Modal instead
Intellectual honesty first: Modal wins clearly when you're the developer with bursty workloads —
- GPU burstsYour agent needs serious accelerators for bounded jobs. Modal's GPU menu (up to B300) is in a different league from a CPU box.
- Spiky workloadsScale-to-zero economics beat an always-on machine when utilization is genuinely bursty.
- Already on ModalYour infra is Modal; the sandbox tier keeps everything in one billing relationship.
One watch-out we track: network egress is now metered ($0.04/GiB over the included allowance). Agents that browse, clone, and ship artifacts can move a lot of bytes — model that before you commit.
When spark-vm is the better fit
When the agent isn't a workload but a resident: Muse, Grok Bot, Dots, or whatever you're running — it keeps files for weeks, runs cron jobs overnight, drives a desktop across sessions, and you'd rather pay for the box than meter every CPU-second, gigabyte, and now egress byte. Flat, ownable, auditable.