70 entries · verified 2026-09-24 · schema llm-api.schema.json
Free LLM APIs: 70 verified entries, 58 of which need no credit card, last checked 2026-09-24.
A free LLM API is the starting point for almost every agent project, and it is the category where lists rot fastest: Cerebras removed its no-card free tier in July 2026, Google moved its Pro models behind billing in April, and Half the providers quoted in older guides now require a payment method before the first request.
Every entry below states the free allowance in the vendor's own units, the rate limit, whether a credit card is required, and whether the free tier trains on your prompts. Providers that quietly stopped being free are still listed — marked as no longer free, with the replacement named — because knowing where not to go is half the value.
If you want to start in the next five minutes: Gemini gives the most daily requests without a card, Groq is the fastest, and Ollama is the only option with genuinely no limit at all.
showing 70 of 70
Jamba family models with a recurring free allowance aimed at evaluation.
✅ no credit cardDoes not train on prompts
One of the few providers whose free credit refreshes on a schedule rather than once.
Qwen family access with a free per-model token grant — the cleanest route to Qwen3 Coder for free.
✅ no credit cardDoes not train on prompts
Grants expire 90 days after activation and do not renew — good for a build sprint, not a standing tier.
Bedrock-hosted Claude, Llama and Titan models on the AWS free tier for new accounts.
💳 card requiredDoes not train on prompts
Card required at signup. Bedrock model access must be requested per model in the console.
No longer free — kept so you do not waste a signup.
No public free API tier; the OSS program grants Claude Max to qualifying maintainers instead.
💳 card requiredDoes not train on prompts
Free chat on claude.ai is separate from API access — there is no free API key.
Speech-to-text plus audio intelligence (summarisation, topic detection) on a no-card starter credit.
✅ no credit cardDoes not train on prompts
Azure-hosted OpenAI, Phi and Llama models with the new Azure free-account credit bundle.
💳 card requiredDoes not train on prompts
Identity verification with a card is mandatory; 30-day clock makes it a sprint budget.
Speech-to-text and neural TTS free allowances inside the Azure free account.
💳 card requiredDoes not train on prompts
ERNIE Speed and ERNIE Lite are permanently free on Baidu's Qianfan console for Chinese-language agents.
✅ no credit cardDoes not train on prompts
Requires a mainland-China verified phone number for full access.
India's national language platform — free STT, TTS, translation and OCR APIs for 22 scheduled languages.
✅ no credit cardDoes not train on prompts
Requires registration approval and a pipeline configuration — the signup is heavier than a typical vendor, but the API is genuinely free.
Wafer-scale inference — still the fastest tokens/sec available — but the permanent free tier ended in July 2026.
💳 card requiredDoes not train on prompts
Downgraded from a no-card permanent free tier to a card-gated 30-day trial. Context is capped at 8K on trial.
Decentralised open-model inference with a community tier that costs nothing for low-volume agent calls.
✅ no credit cardDoes not train on prompts
Community capacity is best-effort; expect queueing at peak. Fine for batch, risky for user-facing agents.
Free caching, rate-limiting and logging proxy placed in front of any provider, including your own free keys.
✅ no credit cardDoes not train on prompts
The single highest-leverage free tool in this list: response caching can stretch a 1,000 req/day quota several-fold.
Serverless inference that runs inside a Worker — no egress cost, no cold-start management, free daily budget.
✅ no credit cardDoes not train on prompts
Budget is denominated in Neurons, not tokens — per-model conversion varies, so budget by testing one call.
Every Cohere model including Command A and the Embed/Rerank family on a free trial key, non-commercial use only.
✅ no credit cardNon-commercial use onlyDoes not train on prompts
Trial keys explicitly prohibit production/commercial use — fine for a POC, not for a shipped product.
Anonymous open-model inference on a decentralised GPU network — 100M tokens per key with no account, no card and no signup, on an OpenAI-compatible endpoint.
✅ no credit cardDoes not train on prompts
The 100M allowance is per key, not per day — issuance is unlimited, so you can mint another key when one is spent. Anonymous keys come from POST /tokens. Direct curl calls frequently trip a Cloudflare challenge (403); the browser chat and the documented endpoints work. Runs on the Gonka compute network.
Low-latency STT and TTS built for voice agents, with a large one-time credit and no card on signup.
✅ no credit cardDoes not train on prompts
Credit is one-time; voice agents burn it fast. Pair with a local Whisper fallback.
Cheap open-model inference with a small signup credit and no card requirement for the trial.
✅ no credit cardDoes not train on prompts
No longer free — kept so you do not waste a signup.
Extremely cheap per-token pricing, but no standing free tier and no signup grant since 2026.
💳 card requiredDoes not train on prompts
DeepSeek weights are open — run V4-Flash locally for free rather than paying the API.
Best-in-class TTS voices with a genuine monthly free character allowance, no card required.
✅ no credit cardNon-commercial use onlyDoes not train on prompts
Free-tier output requires attribution and cannot be used commercially.
Lightweight ONNX embedding library — CPU-friendly local vectors that pair directly with Qdrant.
✅ no credit cardDoes not train on prompts
CTranslate2 reimplementation of Whisper that transcribes locally far faster than the reference build.
✅ no credit cardDoes not train on prompts
Fast open-model inference with a small no-card starter credit.
✅ no credit cardDoes not train on prompts
The account suspends when the $1 is spent. Treat as a smoke test, not a tier.
Frontier and open models behind a GitHub PAT — the lowest-friction free tier if you already have a GitHub account.
✅ no credit cardDoes not train on prompts
Intended for prototyping, not production traffic. Rate limits are per-model.
Production speech APIs with a standing free monthly allowance for STT and WaveNet/Neural2 voices.
💳 card requiredDoes not train on prompts
The STT/TTS free allowance does not expire, unlike the $300 signup credit — this is one of the best long-term free voice stacks.
Vertex AI access to Gemini and Model Garden on new-account credits — the enterprise-labelled alternative to AI Studio.
💳 card requiredDoes not train on prompts
Requires a billing account with a card, so it fails the strictest no-card filter — but the $300 covers a serious prototype.
Frontier multimodal models on a permanent free tier — Flash and Flash-Lite families only since April 2026.
✅ no credit cardTrains on your prompts
Free-tier prompts are used to improve Google products outside the EU/UK/EEA. Pro models left the free tier in April 2026.
Consumer desktop app for running local models on modest hardware, with Python bindings for agents.
✅ no credit cardDoes not train on prompts
Open-weight models at 300+ tokens/sec on LPU hardware, with the most generous request-per-day allowance of any free tier.
✅ no credit cardDoes not train on prompts
The 200K tokens/day ceiling usually binds before the request count does. Llama models were removed from the free plan in 2026.
Free OpenAI-compatible access to Gemini, GPT-5.2, Kimi K2 and 30+ other models for Hack Clubbers — the most generous free tier aimed at teenage developers.
✅ no credit card
Gated on a Hack Club account, which is aimed at students and teenage hackers. Not a general-purpose public tier — do not list it as one.
One router in front of 18 inference partners, with monthly free credits on any account.
✅ no credit cardDoes not train on prompts
The $0.10/month allowance is genuinely small — treat this as a smoke-test credit, not a workload budget.
Open-source offline ChatGPT alternative that can also expose a local OpenAI-compatible API.
✅ no credit cardDoes not train on prompts
Multilingual embeddings, reranker and reader APIs with a large free token allowance on one key.
✅ no credit cardDoes not train on prompts
Single-binary llama.cpp wrapper with a built-in web UI and an OpenAI-compatible endpoint.
✅ no credit cardDoes not train on prompts
Image generation API with a daily free token allowance tied to a free account.
✅ no credit cardDoes not train on prompts
Portable C/C++ inference engine that runs GGUF-quantised models on CPU, GPU or Apple Silicon.
✅ no credit cardDoes not train on prompts
Single-file executable that bundles a model and a server — portable agent runtimes with zero install.
✅ no credit cardDoes not train on prompts
Gateway you can call with the literal placeholder key "unused" — no account, no email, no card, just a base URL and an OpenAI SDK.
✅ no credit card
The widely-repeated "2 req/s, 20 RPM, 100 req/hr" figures do not match the vendor documentation: anonymous access is 10 req/min and a free token from token.llm7.io raises it to 40 req/min. Anonymous traffic only reaches the turbo model set.
Desktop GUI for local models with a one-click local server and built-in model discovery.
✅ no credit cardDoes not train on prompts
Drop-in OpenAI API replacement that also speaks image, audio and embedding endpoints on modest hardware.
✅ no credit cardDoes not train on prompts
Code-specialised completions and fill-in-the-middle model inside the Mistral free Experiment tier.
✅ no credit cardTrains on your prompts
European-hosted Mistral family with a free Experiment mode worth roughly $10/month of credits, no card required.
✅ no credit cardTrains on your prompts
Experiment mode trains on your prompts by default. The 2 RPM limit makes it a batch/offline worker, not an interactive backend.
Apple-silicon-optimised inference and fine-tuning library — the fastest local option on a Mac.
✅ no credit cardDoes not train on prompts
Kimi K2 family with signup credit for long-context Chinese and English workloads.
✅ no credit cardDoes not train on prompts
Free hosted endpoints for NVIDIA-curated open models (Nemotron, Llama, DeepSeek), built for prototyping.
✅ no credit cardDoes not train on prompts
Credits are one-time and not refilled; some models are marked as evaluation-use only.
The default way to run open models locally — one command to pull, serve and expose an OpenAI-compatible endpoint.
✅ no credit cardDoes not train on prompts
No rate limit and no data leaving your machine — the correct answer to 'what is the truly free LLM API'.
No longer free — kept so you do not waste a signup.
No standing free tier. Data-sharing evaluation tokens are the only free path.
💳 card requiredDoes not train on prompts
Unless your org is eligible for the data-sharing token grant, treat OpenAI as paid-only.
One OpenAI-compatible key fronting 19+ `:free` models across providers — the best way to A/B models without multiple signups.
✅ no credit cardDoes not train on prompts
The `:free` roster rotates monthly. Upstream providers frequently return 429s on free variants.
EU-hosted, GDPR-friendly inference over 20+ open-weight models at $0, usable without an API key on the anonymous tier.
✅ no credit card
The API host is oai.endpoints.kepler.ai.cloud.ovh.net — posting to the marketing domain returns 405. Anonymous calls are accepted and then answered with "API rate limit exceeded", which is how the ~2 req/min cap presents itself; no key is required to reach that point. The published catalogue returns pricing of 0 for prompt, completion and request.
No longer free — kept so you do not waste a signup.
Search-grounded Sonar models are paid API only; the free consumer app is not API access.
💳 card requiredDoes not train on prompts
Tavily, Exa and Brave all have free search tiers — see the MCP list for agent-ready alternatives.
Fast, small neural TTS that runs on a Raspberry Pi — the free voice layer for offline agents.
✅ no credit cardDoes not train on prompts
Completely keyless text and image generation API — the fastest possible zero-signup smoke test.
✅ no credit cardDoes not train on prompts
No SLA and no guarantee of model availability — perfect for demos, unsuitable for production.
Client-side JavaScript SDK that gives a web app AI, storage, database and auth with no backend — the developer pays nothing because each signed-in user spends credits from their own Puter account.
✅ no credit card
Usually marketed as "free and unlimited", which is only half true. It is genuinely free for the developer, but not unlimited: once a user exhausts their Puter allocation they pay Puter directly, so an app with heavy anonymous traffic will hit a wall. The open-source developer tools that wrap it as a drop-in OpenAI server are unmaintained third-party bridges and are not listed.
Very fast hosted open-model inference with a permanent per-model daily token allowance.
✅ no credit cardDoes not train on prompts
Quota is per model, so stacking three models triples your effective daily allowance.
Indian-language LLMs (Sarvam-M), speech and translation APIs with signup credits for Indian developers.
✅ no credit cardDoes not train on prompts
Indic-language models are the differentiator; English performance is below frontier tiers.
The standard library for local embeddings and reranking — 15K+ models, no API key, no quota.
✅ no credit cardDoes not train on prompts
Chinese inference host that keeps a rotating set of open models free to call, including embedding and rerank models.
✅ no credit cardDoes not train on prompts
Free model list changes without notice; the reranker and embedding models are the durable part.
Anonymous stealth model on OpenRouter with a 1M-token context window, free at $0 in and $0 out while the provider stays cloaked.
✅ no credit card
Confirmed live against OpenRouter's model API on 2026-09-24. Stealth previews are temporary by design: Ox Alpha, the previous drop, was unmasked as Z.AI's GLM-5.3-Flash and removed from the free catalogue within a week — which is exactly why it is not listed here. Needs an OpenRouter account, and prompts are handled by an anonymous provider, so keep secrets out of it.
Accurate multilingual transcription with a recurring free monthly hour allowance.
✅ no credit cardDoes not train on prompts
Stable Diffusion and Stable Audio via API, with a small recurring monthly credit grant.
✅ no credit cardDoes not train on prompts
Feature-rich local web UI with OpenAI-compatible API, extensions and multi-model loading.
✅ no credit cardDoes not train on prompts
No longer free — kept so you do not waste a signup.
200+ open models, but the signup trial credit was retired; a $5 minimum purchase is required.
💳 card requiredDoes not train on prompts
Hugging Face pipelines running in the browser or Node via ONNX runtime, with WebGPU acceleration.
✅ no credit cardDoes not train on prompts
One endpoint in front of every major provider with free monthly included credit and per-request observability.
✅ no credit cardDoes not train on prompts
Also works as a BYOK proxy — you can route your existing free keys through it for logging and failover.
High-throughput serving engine with PagedAttention — the standard for self-hosted production inference.
✅ no credit cardDoes not train on prompts
High-quality embeddings and rerankers with an unusually large free token grant for RAG prototypes.
✅ no credit cardDoes not train on prompts
The 3 RPM limit is the real constraint — fine for indexing a corpus offline, painful for live retrieval.
Runs quantised LLMs in the browser on WebGPU — zero server, zero cost, zero data egress.
✅ no credit cardDoes not train on prompts
First load downloads 1-2GB of weights into the browser cache — design for that or you will lose users.
llama.cpp compiled to WebAssembly with WASM SIMD and threads — GGUF models straight in the browser.
✅ no credit cardDoes not train on prompts
No longer free — kept so you do not waste a signup.
No documented free tier; console access requires prepaid credits.
💳 card requiredDoes not train on prompts
GLM-4.x Flash models served free of charge, popular as a drop-in for reasoning and coding agents.
✅ no credit cardDoes not train on prompts
Flash tiers are free but rate-limited during peak hours in Asia.
Unified gateway over 100+ models that keeps a small set of model ids permanently free, including DeepSeek V4 Flash Vision and GLM-5.3.
✅ no credit card
The gateway as a whole is commercial — only the -free suffixed model ids cost nothing. Paid tiers advertise a 5% service-fee discount on top-up, so expect upsell copy on the pricing page.
70 free LLM API providers are listed in this catalogue as of 2026-09-24, and 58 of them can be started without a credit card. Each entry records the free limit in the provider's own units, the rate limit, and the date a human last confirmed it.
58 of the 70 providers here can be used without entering card details. The catalogue flags every provider that requires a card and can filter them out in one click, which matters because a card-required free tier is a trial with extra steps.
They are free up to a published quota, not unmetered. Every provider listed states a limit — requests per minute, tokens per day, or a credit balance — and this catalogue records that limit instead of repeating a marketing claim. No legitimate unmetered frontier API exists, so none is listed.
Three catches recur: your prompts may be used for training, commercial use may be restricted, and a free tier can be withdrawn with little notice. All three are recorded per provider here, which is why the entries carry a training flag, a commercial-use flag and a last-verified date.