freeinferencingBy Hackers In the Loop

Free model APIs.
Prototype freely.

Find hosted LLMs, embeddings, speech, image, and research APIs you can actually use, clearly separated from trials and marketing fluff.

Explore the catalog How we verify
Verified Sep 19, 2026 · 57 ongoing · 113 tracked
Current snapshotYAML → static site
Ongoing or recurring
57
Community / conditional
7
Trials / promotions
49
Current offers
113
Under review
38
No free offer found
47

Every classification links back to current provider documentation, pricing, or a live catalog.

Useful before your first invoice

High-confidence, recurring offers with practical model access. Limits still apply.

GroqCloud

Always-free quota
At least $94.55/moEquivalent free value

Google AI Studio / Gemini Developer API

Always-free quota
Up to $1,235.17/moEquivalent free value

Cloudflare Workers AI

Always-free quota
$3.35/moEquivalent free value

OpenRouter

Free model routes
Up to $2,812.60/moEquivalent free value

Free does not mean private.

A provider may retain prompts, permit human review, use free-tier content to improve products or models, or route requests through another company. Those rules can differ between free, paid, and enterprise plans.

How we audit data agreements

Find your free inference route

Search provider names and exact model IDs, then narrow by the kind of free access you need.

a recurring allowance the provider funds indefinitely.specific model routes currently listed at zero price; the set rotates.credits that refresh on a published schedule.no-cost access for research, evaluation, and development, not production use.a one-time allowance for new accounts; it does not recur.community-supplied or end-user-funded capacity with no guaranteed throughput.

113 current providers

Recommended weighs offer durability, verification, privacy, measured free value, and audited popularity. Unmeasured allowances and providers with no traceable adoption rank lower.

GroqCloud

Always-free quota · Private prompts · 13 models

At least $94.55/mo

OpenRouter

Free model routes · Partially private prompts · 25 models

Up to $2,812.60/mo

AwanLLM

Always-free quota · Private prompts · 6 models

At least $79.79/mo

Modal

Monthly credits · Private prompts · 1 model

$30.00/mo

Arli AI

Always-free quota · Private prompts · dynamic catalog

At least $2.33/mo

Cohere

Research / development · Partially private prompts · 10 models

Up to $350.72/mo

Albert API

Research / development · Private prompts · 10 models

At least $6.62/mo

Sail Research

Monthly credits · Private prompts · 12 models

$5.00/mo

Z.AI Model API

Free model routes · Private prompts · 3 models

Not priced

Api.Airforce

Free model routes · Partially private prompts · 28 models

Up to $456.30/mo

Arnict

Monthly credits · Partially private prompts · 2 models

$5.00/mo

ElevenLabs API

Always-free quota · Not private prompts · 5 models

Up to $1.00/mo

Ollama Cloud

Monthly credits · Partially private prompts · 20 models

Not priced

LLM7

Always-free quota · Partially private prompts · 2 models

Not priced

BazaarLink

Free model routes · Partially private prompts · 3 models

Not priced

Cartesia

Always-free quota · Not private prompts · 3 models

Up to $1.00/mo

OrcaRouter

Free model routes · Partially private prompts · 4 models

Not priced

OpenTyphoon API

Research / development · Partially private prompts · 4 models

Not priced

Beam

Monthly credits · Not private prompts · dynamic catalog

$30.00/mo

Kilo AI Gateway

Free model routes · Not private prompts · 22 models

Not priced

LLM.API

Free model routes · Partially private prompts · 1 model

Not priced

DreamPrompting

Always-free quota · Not private prompts · 7 models

Up to $152.19/mo

AI Horde

Community capacity · Partially private prompts · 26 models

Not priced

Deepgram

One-time trial · Partially private prompts · dynamic catalog

$200.00 once

Fikra API

One-time trial · Private prompts · 3 models

$0.15 once

Novita AI

One-time trial · Private prompts · 7 models

Not priced

Palabra.ai API

One-time trial · Private prompts · dynamic catalog

$50.00 once

Pollinations.ai

Free model routes · Partially private prompts · 21 models

Not priced

Tencent Hunyuan

One-time trial · Partially private prompts · 9 models

Up to CN¥10.30 once

BytePlus ModelArk

One-time trial · Partially private prompts · 7 models

At least $2.40 once

ch.at

Community capacity · Partially private prompts · 1 model

Not priced

Speechmatics

One-time trial · Partially private prompts · dynamic catalog

$100.00 once

Waterfall

Free model routes · Partially private prompts · 27 models

Not priced

AI21 Studio

One-time trial · Partially private prompts · dynamic catalog

$10.00 once

Fireworks AI

One-time trial · Partially private prompts · dynamic catalog

$1.00 once

Hyperbolic

One-time trial · Partially private prompts · dynamic catalog

$1.00 once

AI.cc / AICC

One-time trial · Partially private prompts · 3 models

$1.00 once

Baidu Qianfan

One-time trial · Partially private prompts · dynamic catalog

CN¥20.00 once

Puter.js AI

User pays · Partially private prompts · 32 models

Not priced

WaveSpeedAI

One-time trial · Partially private prompts · dynamic catalog

$1.00 once

Amazon Bedrock

One-time trial · Partially private prompts · dynamic catalog

At least $100.00 once

Replicate

One-time trial · Partially private prompts · dynamic catalog

Not priced

Inception Platform

One-time trial · Partially private prompts · 2 models

Up to $75.00 once

SiliconFlow Global

One-time trial · Partially private prompts · dynamic catalog

$1.00 once

Mancer AI

Free model routes · Not private prompts · 1 model

Not priced

Sarvam AI

One-time trial · Not private prompts · 5 models

₹100.00 once

Unbiased Pareto

One-time trial · Partially private prompts · 1 model

Not priced

Logfare

Always-free quota · Not private prompts · 23 models

Not priced

FastRouter

Free model routes · Not private prompts · 20 models

Not priced

Viggle API

One-time trial · Not private prompts · 2 models

$1.00 once

Voyage AI

One-time trial · Not private prompts · 6 models

At least $114.00 once

AssemblyAI

One-time trial · Not private prompts · dynamic catalog

$50.00 once

Ramp Router

One-time trial · Not private prompts · dynamic catalog

$26.00 once

Entrim.ai

One-time trial · Partially private prompts · 11 models

$10.00 once

Upstage

One-time trial · Not private prompts · dynamic catalog

At least $10.00 once

Clarifai

One-time trial · Not private prompts · dynamic catalog

$5.00 once

OpenCode Zen

Limited promotion · Not private prompts · 7 models

Not priced

SEA-LION API

One-time trial · Not private prompts · 5 models

Not priced

Paxa Labs API

One-time trial · Not private prompts · 4 models

$0.10 once

Inferon

One-time trial · Not private prompts · 8 models

Not priced

How to choose a free inference API

A useful free endpoint is more than a zero-dollar price. The offer has to match your model, traffic, data-handling, and deployment needs, and the evidence has to survive past the signup page.

Separate recurring access from trials

An always-free quota or refreshing credit can support repeated development. A signup credit, expiring promotion, or fixed trial window cannot. This catalog keeps them together for discovery but labels them separately, so “current” never silently becomes “permanently free.”

Confirm the exact model and API shape

Provider names are not enough. Check the model ID you will send, whether it is still in the zero-price catalog, and whether the endpoint is OpenAI-compatible or provider-specific. Rotating routers can change their free pool without changing their marketing page.

Read the operational limits

Requests per minute, tokens per day, concurrency, geography, account eligibility, and payment-card rules decide whether an offer works in practice. A generous token allowance can still be unusable for a bursty workload if its request or concurrency ceiling is low.

Treat prompt privacy as a separate decision

Free-tier content may be retained, reviewed, routed to another operator, or used for model training and broader product improvement. “No training” alone does not answer every data-handling question. Each current record shows a conservative privacy classification and the reason behind it.

Compare allowance, not marketing size

Equivalent paid value converts a documented free allowance using the same provider’s published paid rates. It makes unlike quotas easier to compare, but it is not cash, guaranteed savings, or a substitute for latency, reliability, model quality, and policy fit.

Re-check before production

Free models, quotas, and legal terms are volatile. Every record carries a dated snapshot and first-party sources so you can confirm the current state. If the decisive limit lives only behind a login, the catalog says so instead of inventing a number.

Convert quotas into usable throughput

Translate every published limit into the same workload unit before comparing providers. Estimate requests per minute at peak, prompt and output tokens per request, concurrency, retries, and any daily or monthly ceiling. The practical allowance is set by the first limit your workload reaches, not the largest number on the pricing page. A high daily token pool may still be a poor fit for a bursty chat application, while a small recurring quota can be excellent for scheduled extraction or evaluation jobs. Keep trial credit separate from recurring capacity so a successful prototype does not create a surprise migration deadline.

Verify authentication and billing behavior

Check whether the route needs an account, API key, organization, research approval, region, or payment method. “No card required” is useful evidence, but it does not guarantee anonymous access or protect an account from future billing changes. When a card or prepaid balance is present, configure provider-side budgets and alerts before testing. Run one request with the exact account type you plan to use, then inspect the usage dashboard or response metadata. A public catalog proves that a model is advertised; only an authenticated request and its resulting balance can show how your workspace actually behaves.

Exercise the failure path deliberately

Send an invalid model ID, approach a documented rate limit, and observe an exhausted allowance in a safe environment when possible. The application should distinguish authentication errors, quota exhaustion, rate limiting, model removal, provider outages, and malformed requests. Retries need backoff and a hard ceiling so a temporary free-tier limit does not become a retry storm. If a gateway supports automatic fallbacks, confirm that it cannot silently choose a paid route or a model with different data terms. Record the status codes and response fields that your monitoring and user-facing errors will rely on.

Measure quality, latency, and geography

A free allowance is valuable only if the route can do the job. Evaluate representative prompts, long-context behavior, structured output, streaming, tool calls, and safety constraints with a small versioned test set. Measure first-token and end-to-end latency from the region where the application runs, not from a provider demonstration. Community and experimental services may vary more by time of day or capacity. Save the exact model ID and response model field with results because aliases and rotating free pools can change the underlying model without changing your client configuration.

Keep the integration portable

OpenAI-compatible syntax lowers switching cost, but it does not make endpoints interchangeable. Providers can differ in supported parameters, token accounting, streaming events, tool schemas, error bodies, context windows, moderation, and retention rules. Put provider-specific authentication, request options, and error normalization behind a narrow adapter. Keep model selection and base URLs in configuration, and test at least one fallback before it is needed. Portability is especially important for zero-price routes because model catalogs, quotas, and experimental programs can change faster than ordinary paid contracts.

Record the decision and its expiry

When a route becomes a dependency, save the source links, snapshot date, offer class, exact model, limits, privacy rationale, account conditions, and person responsible for re-checking it. Set a review interval based on the cost of failure: a weekend prototype can tolerate more uncertainty than a customer-facing workflow or a pipeline handling sensitive data. Treat a provider changelog, catalog diff, unexpected billing movement, or policy revision as a trigger for an immediate review. This turns “we found a free API” into an auditable engineering decision with a clear owner and exit path.

Use the directories as part of the search

A provider missing from the current list is not automatically overlooked. Search the watchlist for plausible offers whose quota, eligibility, billing, or callability remains unresolved, then check the no-free directory for retired tiers, paid-only APIs, consumer chats, BYOK gateways, and self-hosted models that have been repeatedly mistaken for provider-funded inference. Those negative and conditional records preserve the investigation behind the headline list. They also show exactly what new first-party evidence would change the classification, making the catalog useful for discovery even when the answer is “not verified yet” or “not a free hosted API.”

Use the glossary for exact definitions, read how evidence is ranked, or inspect the complete source dataset. For a fast evaluation, shortlist two routes with different operators, run the same versioned prompt set through both, and keep the evidence and failure behavior beside the quality result.

“Start free” is not always free inference.

We keep disputed offers on a watchlist and document providers whose free tier is retired, UI-only, BYOK, or simply paid.

Open watchlist Visit the shame page