Free access to AI models is not one thing. It’s useful to separate four distinct offers that get lumped together under “free AI”:
- Free API tier — a provider gives you a recurring, limited quota of requests or tokens at no charge.
- Free hosted inference — the model runs on the provider’s infrastructure, gated by quotas, queueing, or account conditions rather than a bill.
- Open-weight / local inference — model weights are available to run yourself. That says nothing about whether a hosted API for the same model is free.
- Trial or promotional credits — temporary access that ends when the credits or the trial window run out.
A provider “having a free tier” can mean any of these, and the practical difference — whether your integration still works next quarter — depends on which one it actually is.
Who’s currently offering what
Terms change, so treat this as a starting point for your own check, not a standing guarantee:
| Provider | Offer type | Worth checking |
|---|---|---|
| Google Gemini API | Free tier via Google AI Studio | Quotas vary by model and project; data-handling terms can differ between free and paid tiers. |
| Groq | Free tier for selected models | Limits may combine RPM, RPD, and TPM; varies by model and account. |
| OpenRouter | Aggregator with models tagged :free | One API across many models and routes; free-tier limits can apply at the account level. |
| Mistral | Free or limited developer tier for select products | Check eligibility and permitted use case per model. |
| Cloudflare Workers AI | Free daily compute allowance, billed beyond it | The allowance is measured in Neurons, not a fixed request count. |
| NVIDIA API Catalog / NIM | Free access for development and evaluation | Not a production SLA or a permanent free allocation. |
| Cohere | Trial/evaluation access | Particularly relevant for embeddings, reranking, and RAG; evaluation terms differ from production pricing. |
| Hugging Face Inference Providers | Aggregated access with free credits | Credits can be small and change; open weights don’t imply free hosted inference. |
| Ollama | Local inference always free; cloud models under evolving terms | See the FAQ below — cloud pricing has been shifting during 2026. |
Why providers offer free access at all
A few recurring reasons: lowering the barrier so developers build on a given API before they’re paying customers; driving adoption of a broader cloud platform (Google, NVIDIA, Cloudflare all sell more than tokens); using otherwise-idle compute capacity; and competing to be the place developers go for a popular open-weight model, since the same model is often servable by several providers at once.
None of that makes inference free for the provider — accelerators, power, and networking still cost money — so a free tier is best read as a customer-acquisition mechanic with a typical funnel: experiment for free, integrate, grow usage, hit the quota, move to a paid plan.
A live example: Ollama’s shift on cloud pricing
Ollama has long drawn a clean line between two different products: running a model on your own machine (free, unmetered, and not something a company can take away) and running a model on Ollama’s own cloud infrastructure (compute the company pays for).
During 2026, independent pricing trackers report that line hardening: Ollama’s cloud-hosted models have been moving toward usage-based, credit-funded billing rather than an open-ended free tier, while local execution stays free regardless. We weren’t able to independently confirm every specific detail of this shift — including an exact cutover date or a precise list of affected cloud models — from third-party sources, and pricing pages for every provider in this space change often. If you depend on Ollama’s cloud tier, check docs.ollama.com directly rather than relying on this summary.
The underlying signal is the useful part regardless of the exact date: free local execution and free cloud hosting are two different promises, and a provider can keep one while ending the other.
Patterns worth watching
- Quotas are multi-dimensional. RPM, RPD, TPM, TPD, concurrency, context-window limits, and project-vs-account scoping can all apply at once, often at different levels for the same provider.
- Free models are getting more task-specific. Not every workload needs one large general-purpose model — classification, embeddings, and RAG reranking are increasingly served by smaller, cheaper, sometimes-free specialized models, while a paid model gets reserved for cases that actually need the extra quality or context.
- Data-handling terms often differ by tier. Whether prompts can be used to improve a provider’s models, how long data is retained, and what region it’s processed in can all change between a free and a paid tier at the same provider — worth checking before sending anything sensitive through a free endpoint.
- Free offers change with little notice. A quota can be narrowed, a tier converted to a trial, or a model moved behind a paywall. A list of “free providers” is a snapshot, not a stable reference.
Why this argues for routing across providers, not picking one
Given how often any single free tier’s terms can move, spreading requests across more than one provider — and failing over automatically when one is unavailable or exhausted — is a more resilient default than committing an integration to one vendor’s current free allowance. That’s the practical reason AI Bakery Gateway treats “free-tier-first, multi-provider” as its default routing posture rather than a one-time integration choice: the providers behind it are expected to keep changing their own terms, not the other way around.
Frequently asked
Which providers currently offer a free AI API tier?
Several platforms offer free or limited developer access, including Google Gemini API, Groq, OpenRouter (models tagged :free), Cloudflare Workers AI, and selected tiers from Mistral, NVIDIA, Cohere, and others. The shape of the offer varies — a recurring quota, trial credits, or a model-specific allowance — so check each provider's own pricing page before integrating.
What's the difference between a free API and a free local model?
A free API runs the model on the provider's servers, bound by their quotas and terms. A local (self-hosted) model runs on your own hardware — no provider bill, but you supply the compute, memory, and electricity, and you manage the deployment yourself.
What do RPM, TPM, and RPD mean in rate limits?
RPM is requests per minute, TPM is tokens per minute, RPD is requests per day. Providers also commonly cap tokens per day, concurrent requests, and context size, sometimes at the project level rather than per API key.
Why do free AI API tiers keep changing?
Cloud inference has real, ongoing costs — accelerators, electricity, networking. Providers adjust free tiers over time: narrowing quotas, moving to trial credits, or shifting to pay-as-you-go for compute-heavy models, while sometimes keeping local or lighter-weight use free. Treat any free-tier list as a snapshot, not a permanent guarantee.
Is it true Ollama is ending free cloud model access?
Multiple independent trackers report Ollama moving its cloud (hosted) models toward usage-based, credit-funded billing during 2026, while continuing to keep locally-run models free of any provider charge. We could not independently confirm a specific cutover date or exact model list beyond Ollama's own pricing page, so check docs.ollama.com directly before relying on either.