Providers whose own page says they do not train on your prompts (data_policy.trains_on_prompts: false). Each provider’s own page on this site shows the page its policy rests on (a link, with the day it was read). Read the conditions: some say it only holds unless you opt in, or in a privacy notice rather than a contract. Providers that do not state it are not listed, which never means they train.
12 of 68 tracked providers match.
| Provider | Type | What's free | The catch | Verified |
|---|---|---|---|---|
| Cloudflare Workers AI no cardno phonecommercial OKOpenAI-compat | Ongoing | 10,000 Neurons/day, all account plans | Resets daily at 00:00 UTC; overage on a Workers Paid plan bills at $0.011/1,000 Neurons. A few models (e.g. Kimi K2.6/K2.7-code, GLM-5.2) now require a Workers Paid plan | 2026-10-08 |
| Cerebras cardcommercial OKOpenAI-compat | Trial | $5 in free credits for new accounts, usable across all public models | A verified payment method is required to activate Playground/API access (no charge until you buy credits). Credits expire 30 days after grant; whether any free access persists past expiry is not stated | 2026-10-08 |
| Fireworks AI no cardOpenAI-compat | Trial | $1 trial credit | Fireworks uses prepaid credits. After the $1 credit is exhausted, add a payment method and credits (or enable auto top-up) to continue; account limits can rise with spend | 2026-10-08 |
| Novita AI no cardeval onlyOpenAI-compat | Trial | $100 promotional credits valid for 90 days for eligible Novita Agent Sandbox usage (CPU, RAM, storage), granted after an account-setup survey; Novita does not say they apply to Model API calls; the pricing page also lists a few LLMs as Free (e.g. Ling 3.1 Flash, Apodex 1.1 Mini) | Signup advertises $100 Sandbox credits valid for 90 days; Novita's sandbox pricing doc describes them as credits for "eligible Novita Sandbox usage" and does not say they apply to Model API calls. No credit card required. Novita Terms of Service state that the Site and Marketplace Offerings may not be exploited for any commercial purpose without express prior written permission. | 2026-10-08 |
| Modal card | Ongoing | $30/month recurring credit (Starter plan) | Recurring monthly credit, not a one-time trial. Modal is serverless compute you deploy models on, not a hosted model API. A payment method on file is required to use Modal at all. | 2026-10-08 |
| Scaleway Generative APIs no cardOpenAI-compat | Trial | 1,000,000 tokens free + 60 min Whisper transcription; billing starts at token 1,000,001 | European provider (France). Free allowance is a one-time token bucket, not time-limited. Official rate limits apply once a valid payment method is registered; identity verification raises them. | 2026-10-08 |
| Jina AI no cardno phonecommercial OKOpenAI-compat | Trial | 10M free tokens (one-time) across all models — embeddings, rerankers, classifier; plus a keyless Reader (r.jina.ai) for basic use | The 10M-token balance is a one-time grant that does not replenish; the keyless Reader is genuinely ongoing. Hosted API is commercial-OK and data is not used for training. No card required. | 2026-10-08 |
| Ollama Cloud no phonecommercial OKOpenAI-compat | Ongoing | $0 Free plan: starter amount of cloud usage credits, limited to a smaller set of starter models; buying credits unlocks all cloud models | First-party — Ollama hosts the cloud models. Requires an ollama.com account + API key. The old 5-hour session / weekly limits are gone: usage is now credit-based with a monthly reset; Pro ($20/mo) includes $60 of credits. Per its cloud page, Ollama does not use cloud prompts and responses to train models; the page says nothing about logging or retention. Whether a card is required is not stated on the current pricing page. | 2026-10-08 |
| Moondream Cloud no cardOpenAI-compat | Ongoing | $5/month usage credits in every workspace (Free plan) for the Moondream vision model — caption, query (VQA), detect, point | Recurring $5/month credit, no credit card required. Per the Terms (§5.2), data is not used to train Moondream's generally available models. Commercial terms for the hosted API not specified. OpenAI-compatible endpoint. | 2026-10-08 |
| Speechmatics no card | Trial | Free plan: $100 credit grant on signup (no payment card) for Speech-to-Text; Text-to-Speech lists the first 1 million characters free | One-time $100 credit grant, no card required; to go beyond it, add a card and move to Pro (pay-as-you-go). Accounts created before 1 August 2026 received a one-time $25 transition credit. No data logging for model training by default. Commercial-use permission not explicitly stated. | 2026-10-08 |
| Rime no card | Trial | Free minutes for every new account (Starter plan): the pricing card says ~800 minutes (~800k characters), while the FAQ and docs say 3,000 | One-time allotment on signup; no credit card required (stated on the pricing page). The pricing table shows ~800 free minutes (~800k characters) while the page’s own FAQ and the docs still say 3,000 — an internal inconsistency. Commercial-use terms not stated. | 2026-10-08 |
| Smallest.ai (Waves) | Trial | $10 in free credits with full access on Pay As You Go — TTS (Lightning v3.1), STT (Pulse) and voice agents; speech-to-speech (Hydra) is beta/request-access | $10 starting credit on Pay As You Go. The models pricing page lists 15 concurrent TTS streams while the API reference still says 1 active TTS request per account — inconsistent. The models table shows Voice Cloning 'No' for Pay As You Go although the TTS card advertises instant cloning. Card and commercial terms not stated. /pricing now redirects to /pricing/agents. | 2026-10-08 |
FAQ
Which free LLM APIs say they do not train on my prompts?
The providers on this page, each according to its own terms or privacy page, cited with the day it was read on the provider’s page on this site. The wording differs: for some it holds only while you have not opted in to training, and a privacy notice is not a contract.
Does a provider missing from this list train on my data?
Not necessarily. A provider is missing when its own page says it may train on prompts, says nothing we could cite, or could not be read. Check the provider before you send sensitive data.