# Hugging Face Jobs (Sandbox API)

Official pricing: https://huggingface.co/docs/hub/en/jobs-pricing  
Category: paas · Isolation: vm

## Pricing regimes (raw)

- **On-demand** (sizes): cpu-basic 2 vCPU/16 GiB $0.01/h; cpu-upgrade 8 vCPU/32 GiB $0.03/h; cpu-xl 16 vCPU/124 GiB $1/h; cpu-performance 32 vCPU/256 GiB $1.9/h [alt]
- **GPU: Nvidia T4** (sizes): t4-small 4 vCPU/15 GiB $0/h; t4-medium 8 vCPU/30 GiB $0.2/h [alt]
- **GPU: Nvidia L4** (sizes): l4x1 8 vCPU/30 GiB $0/h [alt]
- **GPU: Nvidia L40S** (sizes): l40sx1 8 vCPU/62 GiB $0/h [alt]
- **GPU: Nvidia A10G** (sizes): a10g-small 4 vCPU/15 GiB $0/h; a10g-large 12 vCPU/46 GiB $0.5/h [alt]
- **GPU: Nvidia A100 80GB** (sizes): a100-large 12 vCPU/142 GiB $0/h [alt]
- **GPU: Nvidia H200** (sizes): h200 23 vCPU/256 GiB $0/h [alt]
- **GPU: Nvidia RTX PRO 6000** (sizes): rtx-pro-6000 23 vCPU/256 GiB $0/h [alt]

## Features

Yes: persistent disk, what survives a stop, volumes, idle auto-stop, ≥24 h sessions, custom image (Docker or snapshot), your own Docker/OCI image, full VM (own kernel), root, SSH, HTTPS preview URLs, open internet, SOC 2, HIPAA, SSO, GPU, Python, webhooks, scheduled runs (cron), EU data residency, spend limits, gVisor or VM (no shared kernel), inbound access rules, extra volumes, shared volumes, FUSE mounts, kernel features (eBPF, mounts, cgroups)

No: snapshots, memory snapshots, fork/clone, pause/resume, whole disk survives a stop, start from your own snapshot, Docker inside, nested virtualization, browser, desktop GUI, computer-use API, browser + desktop control, anti-bot stealth, CAPTCHA solving, residential IPs, preinstalled agents, public IPv4, egress allowlist, static egress IP, self-hosting, BYOC, open source, arm64, Windows, macOS, Node.js, HTTP method/path egress rules, wake on request, live resize, live fork (no pause), memory fork, MCP server, usage per sandbox (API), automatic snapshots, snapshots on demand, agent harness API, hosted agent API (their own agent), firewall inside (nftables), secret proxy, secret proxy for any API

Unknown: everything else. Evidence (source + quote) per feature: https://battleships.dev/data/providers/huggingface-jobs.json → feature_evidence

## Company

- Ownership: acquired (Nvidia)
- Stage: series-d+ [source](https://techcrunch.com/2023/08/24/hugging-face-raises-235m-from-investors-including-salesforce-and-nvidia/)
- Accelerators: Station F [source](https://stationf.co/programs)
- Total raised: $395M (estimated: Sum of publicly reported rounds (seed, Series A/B/C, and the $235M Series D) as aggregated by Crunchbase/press; no single primary source states a consolidated lifetime total.) [source](https://techcrunch.com/2023/08/24/hugging-face-raises-235m-from-investors-including-salesforce-and-nvidia/)
- Last round: series-d, $235M, 2023-08 (confirmed) [source](https://techcrunch.com/2023/08/24/hugging-face-raises-235m-from-investors-including-salesforce-and-nvidia/)
- Investors: Salesforce Ventures, Google, Amazon, Nvidia, Intel, AMD, Qualcomm, IBM, Sound Ventures, Lux Capital, Sequoia Capital, Coatue, SV Angel
- Revenue: $150M (confirmed: Reported by TechCrunch coverage of the Nvidia acquisition and reflected as company revenue on Wikipedia (2025 figure carried forward).) [source](https://techcrunch.com/2026/09/03/nvidia-confirms-it-will-buy-hugging-face-for-12-9-billion/) ARR, 2026

On Sept 3, 2026 Nvidia confirmed it will acquire Hugging Face for ~$12.93B (~$11.9B cash plus up to $1B in employee retention equity), deal expected to close H1 2027 pending regulatory approval; Hugging Face is to keep its brand and remain an open, hardware-agnostic platform. This card covers 'Hugging Face Jobs', a compute-on-demand product (CPU/GPU containers billed per second) inside the broader Hugging Face Hub/platform, not a separately-funded entity.

## Caveats

- Public Jobs rates, not Spaces free hardware. The huggingface_hub Sandbox API (experimental; formerly hf-sandbox) runs each dedicated sandbox as one HF Job, billed at the Job flavor rate. Jobs bill Starting and Running, per minute; builds free. Exposed ports add $0.01/job-hour.
- Features set to null were not verified. No sandbox was purchased or tested; feature/performance statements are documentation claims. Null maximums are unknown unless a cited note explicitly states unlimited.
- A complete monthly bill requires known snapshot/egress/plan limits; do not silently treat missing ancillary charges as free.
- Re-checked 2026-09-28: jobs-pricing and jobs-overview publish no concurrent-job cap, so plan.concurrency stays null (unknown, not unlimited).
- Sandbox API sandboxes have a fixed 24h maximum lifetime ("The job also has a fixed 24h maximum lifetime as a hard backstop (not configurable)."). max_session_h stays "none" because raw Jobs take any --timeout; sessions over 24h need `hf jobs run`, not the Sandbox API. Source: https://huggingface.co/docs/huggingface_hub/main/en/guides/sandbox
- SandboxPool (Sandbox API) packs many sandboxes into a few shared host Jobs, so per-sandbox cost is the host flavor rate divided by packing density; HF reports 1000 sandboxes created, exec'd and killed in ~16s on 20 cpu-basic hosts (~$0.0009). Pooled sandboxes are uid + Landlock inside one VM, "intended for workloads within the same trust boundary". The engine prices one Job per sandbox, so pooled fan-out costs are overstated. Source: https://huggingface.co/docs/huggingface_hub/main/en/concepts/sandbox

## How this provider charges


As of 2026-09-28. Prices are USD unless explicitly labelled otherwise. Source type: official published list/promo or unknown, never a fabricated quote.

Public Jobs rates, not Spaces free hardware. Harbor hf-sandbox wraps Jobs; an interactive wrapper does not change underlying job billing. Jobs bill Starting and Running, per minute; builds free. Exposed ports add $0.01/job-hour.

| Regime | When it applies | How billed / numbers | Source |
|---|---|---|---|
| CPU Jobs | Starting/Running billed per minute | 2/16 $0.01/h; 8/32 $0.03/h; 16/124 $1/h; 32/256 $1.90/h. | https://huggingface.co/docs/hub/en/jobs-pricing |
| Ports | One or more public exposed ports | $0.01/job-hour extra, not per port. | https://huggingface.co/docs/hub/en/jobs-pricing |
| GPU Jobs | Alternative hardware, not CPU price additive | T4-small $0.40/h, T4-medium $0.60/h, L4 $0.80/h, A10G-small $1/h, A100-80GB $2.50/h, H200 $5/h; whole machine rates. | https://huggingface.co/docs/hub/en/jobs-pricing |

## Gotchas

- Public Jobs rates, not Spaces free hardware. Harbor hf-sandbox wraps Jobs; an interactive wrapper does not change underlying job billing. Jobs bill Starting and Running, per minute; builds free. Exposed ports add $0.01/job-hour.
- Features set to null were not verified. No sandbox was purchased or tested; feature/performance statements are documentation claims. Null maximums are unknown unless a cited note explicitly states unlimited.
- A complete monthly bill requires known snapshot/egress/plan limits; do not silently treat missing ancillary charges as free.

## Worked workload

4 vCPU / 8 GiB; 50 simultaneous instances × 8 h/day × 22 days = **8,800 instance-hours**, **35,200 vCPU-hours**, **70,400 GiB-hours**. At 30% observed CPU:10,560 active-vCPU-hours only where the meter actually uses utilization. Assume one start/instance/day:1,100starts. Retain50GiB snapshots for a month;100GiB outbound. Active disk capacity was not specified.

Cheapest adequate CPU preset is 8 vCPU/32 GB at $0.03/h: $264 for 8,800h. Exposed-port surcharge if required adds $88 → $352. This overprovisions resources; 30% CPU does not discount wall-clock price. Ephemeral jobs do not establish 50GiB retained snapshots; C=50 needs quota validation.

## Sources

- https://huggingface.co/docs/hub/en/jobs-pricing
- https://huggingface.co/docs/hub/en/jobs-overview
- https://pypi.org/project/hf-sandbox/