- The Short Answer: Yes, With Three Kinds of “Free”
- Free Form 1: Permanent Free-Tier LLM APIs (Rate-Limited, Never Expire)
- Free Form 2: Signup-Credit LLM APIs (Capped Amount, Then You Pay)
- Free Form 3: Self-Hosted Open-Source LLMs (No API, No Bill, No Rate Limit)
- Free LLM API Providers Compared
- Which Free LLM API Should You Actually Use?
- Conclusion
- FAQ
Yes — several providers give you a free API key to call open-source LLMs, but “free” splits into three very different deals: a permanent free tier with rate limits that never expires, a one-time signup credit that runs out, or model weights you have to download and host yourself. Below are 6 providers active as of August 2026, sorted by which bucket they fall into, so you can pick based on how long you actually need free access rather than getting surprised by a bill in week three.
The Short Answer: Yes, With Three Kinds of “Free”
If you want a key today and don’t want to enter a card, OpenRouter and Groq both hand out permanent free access with no expiration — you just live with rate limits. If you want higher throughput temporarily, Together AI, Novita AI, and Hugging Face give you free credits or free models to start, then bill you once you’re past that allowance. If you don’t want to depend on any provider’s uptime or pricing changes at all, open-weight models are downloadable straight from the Hugging Face Hub and you can run them on your own hardware — that’s not an API call to Hugging Face’s hosted service, but it’s the only option that’s free forever with zero rate limit.
Free Form 1: Permanent Free-Tier LLM APIs (Rate-Limited, Never Expire)
These don’t ask for payment info and don’t shut off after a trial period. The catch is throughput, not time.
OpenRouter lets you call any model tagged :free in its catalog without adding credits, capped at 20 requests per minute and 50 requests per day. Buy $10 in credits (one time, they don’t expire) and the daily cap rises to 1,000 requests, with the per-minute cap unchanged. As of writing, OpenRouter lists 17 models under the :free suffix, including Z.ai’s GLM-5.2, Google’s Gemma 4 (26B and 31B variants), and several Nvidia Nemotron models. (Checked 2026-08-18 via openrouter.ai/api/v1/models and the platform’s rate-limits documentation at openrouter.ai/docs/api_reference/limits.)
Groq requires no credit card for its free tier and doesn’t gate it behind a trial window, but every model has its own per-minute and per-day request and token caps. For example, openai/gpt-oss-120b is capped at 30 requests/minute, 1,000 requests/day, 8,000 tokens/minute, and 200,000 tokens/day on the free tier — small text models like the Llama Prompt Guard variants get much higher token ceilings (500K tokens/day). Paying customers on the “on-demand” tier get roughly 30-40x the request and token headroom, but the free tier itself doesn’t expire. (Checked 2026-08-18 via Groq’s published rate-limits table.)
Novita AI keeps a rotating set of open-weight models priced at $0 per token as a standing part of its catalog rather than a time-boxed promo — currently Qwen3.5-Plus, Qwen3.6-Plus, and Bunny, each at 0/0 input-output pricing per the live model list. This is separate from Novita’s dedicated free-model roundup covering Llama, Qwen, GLM, and BGE-M3, and separate again from the platform-wide referral program below. Free-tier model selection here rotates as new models launch and older ones move to paid pricing, so check the live catalog rather than assuming a specific model stays free. (Checked 2026-08-18 via api.novita.ai/openai/v1/models.)
Free Form 2: Signup-Credit LLM APIs (Capped Amount, Then You Pay)
This is the most common “free” you’ll run into, and it’s the one that surprises people — the meter is running from day one, it just starts at zero.
Novita AI’s referral program gives $10 in LLM API credits to both the referrer and the new signup, with a stated cap of up to $500 total earnable through referrals. That’s a wallet top-up, not a rate-limited tier — once the $10 is spent, standard per-token pricing applies. (Checked 2026-08-18 via novita.ai/referral.)
Together AI removed its general free trial in a July 2025 billing change — new accounts now need a minimum $5 prepaid balance before making calls. It does, however, run specific model endpoints priced at $0.00/1M tokens outside that prepaid requirement, such as Prism-ML’s Ternary-Bonsai-27B, a 27B open-weight model with a 262K context window. Startups can also apply separately for up to $50,000 in credits through its Startup Accelerator program, which is an application process, not a signup perk. (Checked 2026-08-18 via together.ai/models/prism-ml-ternary-bonsai-27b pricing block and Together’s billing docs.)
Hugging Face’s Inference Providers is a hosted API — you call it remotely with a Hugging Face access token, the same way you’d call OpenRouter or Together AI, and Hugging Face routes the request to the underlying provider and bills you. Every free account gets $0.10 per month in credit automatically, rising to $2.00/month on the $9/month PRO plan. That’s enough for light testing of small models, not production traffic — the credit renews monthly rather than being a one-time grant, but the amount is small enough that most real workloads exhaust it within a handful of requests. (Checked 2026-08-18 via huggingface.co/docs/api-inference/en/pricing.)
DeepInfra advertises “$10 USD free per month” on its pricing page, structured as a recurring monthly allowance rather than a one-time signup bonus. (Checked 2026-08-18 via deepinfra.com/pricing.)
Free Form 3: Self-Hosted Open-Source LLMs (No API, No Bill, No Rate Limit)
This isn’t an API key at all — it’s downloading the model weights yourself and running them on hardware you control, through something like vLLM, Ollama, or llama.cpp. The Hugging Face Hub is the primary distribution point for open-weight releases (Llama, Qwen, GLM, and most other open-source families all publish weights there) — that’s a separate thing from Hugging Face’s own hosted Inference Providers API described above; the Hub just hosts the files. You pay for compute instead of tokens, and there’s no rate limit because there’s no shared service — but you’re responsible for the GPU, the serving stack, and the uptime. This path makes sense if you need guaranteed availability or you’re processing enough volume that a rented GPU beats per-token pricing; it’s the wrong choice if you just want to test a model this afternoon.
Free LLM API Providers Compared
| Provider | Free type | What you get | Expires? |
|---|---|---|---|
| OpenRouter | Permanent free tier | 50 req/day (or 1,000/day after $10 credit purchase), 20 req/min, :free models only | No |
| Groq | Permanent free tier | Per-model RPM/RPD/TPM/TPD caps, e.g. 1,000 req/day on gpt-oss-120b | No |
| Novita AI (free-tier models) | Permanent free tier | Selected open models at $0/M tokens, catalog rotates | No, but model list changes |
| Novita AI (referral) | Signup credit | $10 in credits per referral, cap $500 total | Yes, once spent |
| Together AI | Signup credit + free endpoints | $5 prepaid minimum for general use; specific $0.00/M model endpoints (e.g. Ternary-Bonsai-27B) | Free endpoints: no. General credit: none offered since Jul 2025 |
| Hugging Face (Inference Providers, hosted API) | Signup credit (monthly) | $0.10/month free, $2.00/month on PRO | Renews monthly, small amount |
| DeepInfra | Signup credit (monthly) | $10 USD free per month | Renews monthly |
| Self-hosted (weights from the Hugging Face Hub) | No API | Full model weights, run on your own GPU | Never, but no API/rate-limit benefit either |
Which Free LLM API Should You Actually Use?
If you’re prototyping and want zero setup friction, start with OpenRouter or Groq — no card, no expiration, and you’ll know within an hour whether the rate limit is a real constraint for your use case. If you’re building something that needs steady throughput before you’re ready to pay, a signup credit like Novita AI’s referral $10 or Together AI’s $0.00-priced model endpoints buys you real testing room without the tight per-minute caps that permanent free tiers impose. If you’re already committed to a specific model and need it running indefinitely without depending on any provider’s rate limits or pricing changes, download the weights from the Hugging Face Hub and self-host.
None of this is “free forever with no limits” — that combination doesn’t exist for hosted APIs, because someone is paying for the GPU time underneath. Permanent free tiers cap requests instead of dollars; signup credits cap dollars instead of requests; self-hosting removes both caps but hands you the infrastructure bill and the operational work instead.
Conclusion
Yes — free open-source LLM APIs exist today, but “free” is a label covering three different deals, not one. Permanent free tiers (OpenRouter, Groq, Novita AI’s rotating $0-priced models) never expire but cap your request or token throughput. Signup credits (Novita AI’s referral, Together AI’s prepaid-exempt endpoints, Hugging Face, DeepInfra) give you a real dollar allowance that runs out once you’re past it. Self-hosting open weights from the Hugging Face Hub removes both kinds of caps entirely, at the cost of running and paying for the GPU yourself. Which one is “the free option” for you depends on whether your constraint is rate limits, a spending ceiling, or long-term independence from any provider’s pricing decisions — not on finding a single provider that avoids all three tradeoffs at once, because none currently does.
FAQ
Are open-source LLM APIs actually free, or is there always a catch?
There’s always a catch, and it’s one of three shapes: a request-per-day/minute cap that never expires (OpenRouter, Groq), a dollar credit that runs out (Novita AI referral, Together AI’s $5 minimum, Hugging Face, DeepInfra), or the requirement to self-host on your own hardware. Knowing which shape you’re getting before you build on it avoids the surprise bill or the surprise 429 error.
Which provider has the most generous permanent free tier?
Groq’s free tier gives higher token ceilings on some models — up to 200,000 tokens/day on gpt-oss-120b and 500,000 tokens/day on smaller guard models — compared to OpenRouter’s flat 50 requests/day cap before you add credits. Which one wins for you depends on whether your bottleneck is request count or token volume; a chat app with short exchanges hits OpenRouter’s request cap first, while a batch-summarization workload hits token limits first.
Can I combine free tiers across multiple providers to get more capacity?
Yes, and it’s a common pattern for testing — route different traffic to different providers’ free tiers rather than relying on one. Just be aware that free-tier model selection changes over time (a model that’s :free on OpenRouter today may not be next quarter), so build for provider fallback rather than hardcoding one specific free model ID.
Is Novita AI’s $10 referral credit the same as its free-tier models?
No, they’re separate. The referral program is a one-time $10 credit toward standard per-token billing, capped at $500 total across all referrals. The free-tier models are specific catalog entries priced at $0 per token on an ongoing basis, independent of whether you’ve used a referral code. Check the live model catalog for which models currently carry $0 pricing, since that list rotates as new models launch.
If I self-host an open-source model, is it really free?
The model weights are free to download, but running them isn’t — you’re paying for GPU compute, storage, and your own uptime instead of per-token API pricing. For low, steady traffic this is usually more expensive than a paid API tier once you account for idle GPU time; it becomes worthwhile mainly at high, predictable volume or when you need guaranteed data locality.
