Alpha
All entriesContact
Inference provider

Fireworks AI

HQ
United States
Serves
11 open model families
Pricing
mixed
API
OpenAI-compatible

You substantially own inference here. The binding Terms commit Fireworks not to train on or improve the Service with your Content and to run Zero Data Retention for inference, you own the Content, and the DPA gives a 30-day return and deletion right. The contract has no express confidentiality duty over your Content. Retention exceptions exist for the Response API (set store=false), prompt caching and fine-tuning. Serverless processing is not region-pinned. The sub-processor list and compliance reports sit behind a JavaScript-only trust centre whose content is unverified. The DPA has Schedules 1 to 3 only and does not name Anthropic.

Do you really own it?
Substantial
none·limited·partial·substantial·full
Analytical input: AOI B · 71.6/100
The four ownership factors

Floor-weighted, not averaged. Nothing is weak and both use & modify and data control are strong, so the floor is high; reliability sits at moderate, which is what keeps it short of full.

1

Use and modify freelyCan you use it freely and leave without lock-in?

Strong

An OpenAI-compatible API over a predominantly open-weight catalogue keeps workloads portable: the same checkpoints run elsewhere or self-hosted, independent of Fireworks' own contract terms.

How this scores (AOI sub-dimensions)
Transparency & lock-in4/5how portable it is and how easily you can leaveAn OpenAI-compatible API (Chat Completions and Completions) over a predominantly open-weight catalogue keeps workloads portable.
Cost4/5how the pricing model compares and how predictable it isPer-token serverless pricing (cached-input rate, Priority tier, a 50% batch discount, a 30% spot discount on select models) and per-GPU-second dedicated pricing are all public in Fireworks' own pricing docs and broadly competitive for the class (e.g.
2

TransparencyAre the binding terms published, legible and independently checkable?

Strong

The Terms of Service (Last Updated 2026-07-10), the DPA (v3.4) and the Privacy Notice are published. The Terms and DPA are served as PDFs behind a redirect, so they are legible but not easy to find. The sub-processor list and Trust Center are JavaScript-only with unverified content, and the SOC 2 and ISO attestations are vendor statements with no auditor or certificate date.

How this scores
Not a scored AOI dimension. For a hosted provider, transparency is whether the binding terms are published, legible and were actually read - the read/unverified evidence below, not a certification. A strong rating here must trace to a retrieved binding document.
3

ReliabilityDoes it stay up and stay secure?

Moderate

A public status page shows 22 of 23 endpoints at 100 percent and the lowest at 99.50 percent over 90 days with no listed incidents. The standard Terms carry no SLA or credit clause, and the enterprise 99.9 percent uptime terms are not public.

How this scores (AOI sub-dimensions)
Reliability3/5whether it stays up, with an SLA and status historyA public status page (status.fireworks.ai) shows 23 tracked serverless endpoints with 90-day uptime of 100 percent on 22 and 99.50 percent on the lowest (Kimi K3.1 Pro), and no listed incidents.
Security3/5the controls protecting your traffic and dataDedicated workloads run in "logically isolated environments, preventing cross-customer access or data leakage" per the data-security docs, with TLS 1.2+ in transit, AES-256 at rest, least-privilege access controls and customer-managed keys.
Compliance4/5which independent certifications and attestations it holdsFireworks' docs certifications FAQ, pricing page and blogs announce SOC 2 Type II (2023-10-27), HIPAA compliance (BAA on enterprise) and triple ISO certification (27001/27701/42001, 2025-11-19).
4

Doesn't extract your dataDo the binding terms keep your data and IP yours?

Strong

The binding Terms say Fireworks will not use Content to train or improve the Service and runs Zero Data Retention for inference with no human-review logging (Sec 3.6). The customer owns Content (Sec 3.2) and the licence is limited to providing the Service (Sec 3.4). DPA 11.2 gives a 30-day return and then deletion. The contract names exceptions: the Response API, prompt caching, and training, fine-tuning or agent features. Residency is not pinned for serverless and the sub-processor list is unverified.

How this scores (AOI sub-dimensions)
Data governance4/5retention, training-on-inputs and data ownershipToS Sec 3.6 commits that Fireworks "will not use your Content to train our own models or to improve the Service" and runs Zero Data Retention for inference (no human-review logging, no retention beyond generating and delivering Output).
Residency3/5where your data is processed and storedOn-demand/dedicated deployments can be pinned to US, Canada, Europe or APAC, but "Only GLOBAL is available by default" and quota for any pinned region "must be granted by Fireworks", with a 1.5x price for region-restricted deployments.

How the AOI score is computed

The seven dimensions above, each scored 0 to 5, weighted and summed to the 0 to 100 headline. The score is the analytical input behind the ownership verdict, not the verdict itself.

DimensionScoreWeightPoints
Data governance4/50.2419.2
Compliance4/50.1814.4
Residency3/50.169.6
Security3/50.148.4
Reliability3/50.127.2
Transparency & lock-in4/50.108.0
Cost4/50.064.8
HeadlineB · 71.6/100
Dossier coverageAssess 81%Implement 44%Use 43%Support 57%How complete our four-domain documentation is, a measure of our coverage, not of the model. Each domain links to its page.

Sources

Every rating traces to a primary document. Read means the text was verified; unverified means it is known to exist but has not yet been read.

DocumentWhat it grounds
Terms of serviceread2026-11-19
Terms of Service, Last Updated July 10, 2026 (the page URL redirects with HTTP 308 to a text PDF on cdn.sanity.io).
Privacy Policyread2026-11-19
Fireworks Privacy Notice (states it is not a contract): 'We do not use your prompts, training data, or API inputs to train or improve our AI models without your explicit opt-in'.
Documentationread2026-11-19
Fireworks data-handling docs: ZDR for open models with data held only in volatile memory during the request.
Data Processing Addendumread2026-11-19
Data Processing Addendum v3.4 (the page URL redirects with HTTP 308 to a text PDF).
Subprocessorsunverified2026-11-19
Sub-processor list referenced by DPA 6.2 at trust.fireworks.ai/subprocessors.
Securityread2026-11-19
Fireworks data-security docs: 'Data is encrypted in transit (TLS 1.2+) and at rest (AES-256).' 'Dedicated workloads run in logically isolated environments, preventing cross-customer access or data leakage.' Least-privilege access controls; minimal staff access to production, logged and reviewed.
Vendor announcementread2026-09-20
Fireworks blog post announcing SOC 2 Type II certification and HIPAA compliance (published 2023-10-27, auditor unnamed).
Vendor announcementread2026-09-20
Fireworks blog post announcing triple ISO certification: ISO 27001, ISO 27701, and ISO 42001 (published 2025-11-19, certification body unnamed).
Securityunverified2026-11-19
Fireworks' Trust Center portal at trust.fireworks.ai: a Vanta single-page application that returns HTTP 200 but renders no certificates, reports or sub-processor list without JavaScript.
Documentationread2026-11-19
Fireworks serverless pricing docs and pricing page: per-token pricing (input / cached input / output) with Standard, Priority (1.25x), Fast, Ultra-fast (preview), Batch (50% off) and Spot (30% off select models) modes.
Documentationread2026-11-19
Fireworks on-demand/dedicated-deployment docs: self-serve deployment via firectl/REST/SDKs/console, billed per GPU-second.
Vendor announcementread2026-11-19
Fireworks 'Virtual Cloud' (BYOC) blog post (dated 2025-06-16): the inference engine runs inside the customer's own VPC; access is enterprise, not self-serve.
Vendor announcementread2026-11-19
Fireworks Sovereign product page: 'Fireworks operates the inference engine, and you retain ownership of the environment, cluster, GPUs, and networking.' 'Inference and response handling run inside your cluster.' Offered to Enterprise customers, priced by contract; training is 'coming soon'.
Slaread2026-11-19
Fireworks public status page: 'We're fully operational'; 23 serverless components, 90-day window labelled Aug 2026 to Nov 2026; Kimi K3.1 Pro lowest at 99.50 percent, the other 22 at 100 percent; no incidents listed; no SLA or credit terms.
Documentationread2026-11-19
Fireworks model catalogue page: serverless-tagged families include Kimi K3.1 Pro and K3, DeepSeek V4.1 Flash, GLM 5.3 Flash and GLM-5.3, MiniMax M3.1 Flash and M3, Qwen 3.8 Max, Muse Spark Open, Ember-1, Nemotron Lightning 3.5, Rune 26B-A4B, Ultramarine 1 Flash.
Documentationread2026-11-19
Fireworks OpenAI-compatibility docs: base URL 'https://api.fireworks.ai/inference/v1'; documents Completions and Chat Completions.