Implement · DeepInfra
Can you run it?
Ownership levelPartialnone·limited·partial·substantial·fullAnalytical input C ยท 61.6/100
This page is a projection of the one entry record, the Reliability and Data control factors that Implement covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.
Which domain expands which factor
- AssessUse & modify + Transparency
- ImplementData control + Reliability
- UseReliability
- SupportTransparency
API integration
DeepInfra exposes an OpenAI-compatible API over NVIDIA GPUs (public models run on H100 or A100
per the pricing page), plus per-hour dedicated GPU instances (A100 to B300) for reserved
capacity. An existing OpenAI client integrates by changing the base URL and API key, giving access
to the public catalogue (232 entries across modalities, 93 text-generation). This is a
configuration change, not a rewrite.
Authentication & account setup
API keys go in an Authorization: Bearer header. Each key carries scopes chosen at creation and
editable later; a new key starts with inference, models:read, account:read, lora:read and
voice:read. DeepInfra also issues short-lived, scope-limited JWTs you can hand to third
parties. Organisation-level controls beyond key scopes are unverified; confirm them in the
dashboard if you need them.
Rate limits, tiers & quotas
On public models every account has a default of 200 concurrent requests per model, which you can
ask to raise; the limit is concurrency, not requests per minute. Private dedicated endpoints have
no account rate limit and are bounded by their GPUs, returning 429 Model busy when overloaded.
Per-request service tiers: Standard (best effort at peak), Priority (1.5x price) and Flex (0.8x
price, with occasional unavailability).
Region & ZDR configuration
Zero Data Retention is contractual and applies to every account on DeepInfra-hosted models, so no
toggle is required; retention for batch, asynchronous media or paid cache is opt-in. Region
control is absent: the trust-centre FAQ says US only, no EU region and no choice of US region. If
EU residency is a requirement, DeepInfra does not meet it.
Portability & exit
For the open-weight catalogue the exit path is simple: the OpenAI-compatible endpoint over
portable checkpoints means the same models run on other providers or self-hosted. The exception
is the partner models (94 of 232 public entries, including Anthropic Claude and Google Gemini),
which are not portable and carry the partner's data terms. Older catalogue entries carry a
deprecated timestamp and a replaced_by pointer, and the status page lists scheduled
deprecations, so pin versions you can run elsewhere.
How this scores
The ownership factors this domain covers, drawn from the one entry record.
3
ReliabilityDoes it stay up and stay secure?
ModerateA public status page shows measured 90-day uptime, but there is no uptime SLA or credits and no incident history page, and no independent pen-test report is verified; SOC 2 is Type 1.
How this scores (AOI sub-dimensions)
Reliability3/5whether it stays up, with an SLA and status historyA public status page (status.deepinfra.com) shows measured 90-day uptime for the API and website (100%) and per model (99.98% to 100%), a maintenance page and scheduled deprecations with replacements.
Security3/5the controls protecting your traffic and dataThe trust centre states SOC 2 and ISO 27001 as "Compliant", lists a pen-test report and ISMS manual, and shows control families (MFA, network and endpoint controls, vulnerability remediation).
Compliance3/5which independent certifications and attestations it holdsTerms Sec 5(a) incorporates a DPA by reference that applies to personal data in Customer Data without a separate signature.
4
Doesn't extract your dataDo the binding terms keep your data and IP yours?
ModerateFor DeepInfra-hosted models the Terms make ZDR contractual and controlling, add a no-train clause and a mutual confidentiality duty, and retain customer IP. It falls short of strong because partner models (94 of 232 public entries) are governed by the partner's terms, exceptions allow feature-elected retention, new sub-processors need no consent, residency is US-only with CLOUD Act exposure, and nothing independent attests the ZDR.
How this scores (AOI sub-dimensions)
Data governance4/5retention, training-on-inputs and data ownershipZero Data Retention is contractual and controlling (ToS Sec 7(b), Sec 5(a)).
Residency1/5where your data is processed and storedThe trust-centre FAQ states all inference runs in United States data centres for both the serverless API and dedicated deployments.
What this means for adoptionYou substantially control your data on DeepInfra-hosted open models: the Terms (last modified 2026-10-15) make Zero Data Retention contractual and controlling, forbid training on it, give a mutual confidentiality duty and keep your IP, and a DPA is incorporated by reference. Ownership is capped at partial because that control is not uniform: partner models (Anthropic, Google and 90 more entries) carry the partner's data terms, so route no sensitive data through them. All processing is in the United States with no EU region, so DeepInfra is ruled out where EU data sovereignty is required. The DPA PDF and the attestation documents are unverified, so the SCC content and the SOC 2 and ISO 27001 status rest on the provider's own trust-centre statements.
Sources
The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.
Documentationread2026-11-19
DeepInfra data-privacy documentation: inference input is held in memory only and deleted once inference completes.
Privacy Policyread2026-11-19
DeepInfra Privacy Policy (last modified 2026-08-15) and data-privacy docs say, for inference inputs and outputs: "We will not store, sell, or train using this data unless we have your explicit consent".
Terms of serviceread2026-11-19
DeepInfra Terms of Service (last modified 2026-10-15).
Securityread2026-11-19
DeepInfra trust centre (Sprinto): compliance badges SOC 2 "Compliant", ISO 27001 "Compliant", GDPR "In progress", HIPAA "Compliant".
Vendor announcementread2026-11-19
DeepInfra pricing page (first-party), per million tokens input / output: Llama 3.1 8B Turbo $0.02 / $0.04, Llama 3.3 70B Turbo $0.10 / $0.32.
Securityread2026-11-19
DeepInfra trust-centre FAQ (publisher statement): "All DeepInfra inference runs in data centers in the United States"; "Serverless API and dedicated deployments both run in the US only.
Documentationread2026-11-19
DeepInfra live model listing (GET https://api.deepinfra.com/models/list): 411 entries, of which 232 are public and not deprecated (93 text-generation), and 94 of those 232 carry is_partner (Anthropic 9, Google 13, Bria 20, ByteDance 9 and others).
Terms of serviceread2026-11-19
DeepInfra Terms of Service (last modified 2026-10-15).
Data Processing Addendumunverified2026-11-19
DeepInfra Data Processing Agreement: incorporated by Terms Sec 5(a) and listed on the trust centre as "2026-10-05 - Deep Infra - DPA.docx.pdf".
Subprocessorsread2026-11-19
DeepInfra sub-processor list (docs page stamped last updated 2024-09-06, and the trust centre): Stripe (payment processor), Amazon Web Services and Google Cloud Platform (infrastructure), all United States.
Documentationread2026-11-19
DeepInfra authentication docs: API keys sent as "Authorization: Bearer", each with scopes chosen at creation and editable later (a new key starts with inference, models:read, account:read, lora:read and voice:read), and short-lived scoped JWTs that can be issued to third parties.
Documentationread2026-11-19
DeepInfra rate-limits docs: rate limits apply only to public models; "every account has a default limit of 200 concurrent requests per model"; "Private dedicated endpoints have no rate limits" and return 429 Model busy when overloaded.
Slaread2026-11-19
DeepInfra status page: core services operational, all 172 monitored models operational, API and website 100% 90-day uptime over 86.4 of 90 days measured, per-model figures 99.98% to 100%, a maintenance page, and one scheduled deprecation (embeddinggemma-300m, 2026-11-23) with a named replacement.