Free-LLM: a directory of free LLM APIs with copy-paste setup

What Free-LLM actually solves for a developer choosing an API

The README opens with the problem it targets: finding a free LLM API means "hunting through a dozen changelogs, signing up for five platforms just to compare rate limits, or guessing which provider still has a free tier this month." That is a real cost. Free tiers differ on three axes that matter before you write any code: whether a credit card is required, what the rate and daily limits are, and which models are reachable on the free plan. Free-LLM puts those three in one table.

The audience is narrow but well defined. It suits someone prototyping on a zero budget, a student, or an engineer who wants to point a coding assistant at a free backend rather than a paid one. The README explicitly names Claude Code, Cursor and Codex CLI as tools you can redirect. It does not suit anyone who needs a capacity guarantee, an SLA, or a single vendor to call when something breaks.

The card-requirement column is the most useful editorial decision in the repository. Rows are labelled No, Phone verification, Registration, or similar, so you can filter out providers that would ask for payment details before you even open a signup page.

How the directory is structured and kept current

The repository is a JavaScript project whose main artifact is documentation, not a runtime. The README contains the provider tables, and the code-examples directory holds per-provider snippets plus tool-specific configuration files such as code-examples/claude-code.md, code-examples/cursor.md and code-examples/codex.md.

The tables are generated between marker comments. You can see the pattern in the README: a STATS block wrapped in HTML comments, and provider tables wrapped in comments like TABLE:PERMANENT:START. That layout implies a generator writes content between those markers, which is how the repository keeps a static README aligned with the live site at free-llm.com. The README states the sync happens daily.

The provider list is split by tier type. One section covers permanent free tiers, described as "ongoing free access with rate-limited quotas that never expire." Other sections cover trial credits and no-card options. Each row carries the provider, card requirement, rate limit, daily limit, monthly limit and key models, and links to a page at free-llm.com/provider/<slug>.

One caveat about the numbers. The header in the README says 120+ free LLM models from 41 providers, while the surrounding description mentions 34+ APIs and the README body refers to "all 33 providers" in the code-examples directory. Those figures are inconsistent inside the same document, which is a sign that the counts are maintained by different parts of the pipeline and drift.

Installing nothing: using Free-LLM as a reference and its snippets as code

There is no package to install. The README's three steps are to pick a provider from the directory, get an API key from the provider's key page, then plug the base URL and a model ID into the snippets. Cloning the repository gives you the tables and the code-examples directory locally.

The README's Python example uses the OpenAI SDK against a provider that exposes an OpenAI-compatible endpoint. The README notes that most listed providers do, so the same client works after swapping base_url and api_key.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.groq.com/openai/v1",  # free, no credit card
    api_key="GROQ_API_KEY",                      # get at console.groq.com/keys
)

response = client.chat.completions.create(
    model="llama-3.3-70b-versatile",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(response.choices[0].message.content)
# Groq free tier: 30 RPM, 14,400 requests/day — generous for personal use

The README annotates this block with the Groq free tier as 30 RPM and 14,400 requests per day, and points to console.groq.com/keys for the key. If the request succeeds you get the model's reply printed; if the key is wrong or the quota is exhausted, the SDK raises an error, which is the first place the directory's numbers meet reality.

For coding assistants the README describes environment variables rather than Python. It states that Claude Code reads ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN, and that Codex CLI reads OPENAI_BASE_URL and OPENAI_API_KEY. Set those in the shell that launches the tool, and the tool talks to the free backend instead of its default one. The README defers the details of each tool to its file under code-examples/.

Where the directory model breaks down

A static table cannot enforce anything. The README itself hedges on at least one row, marking Z.AI's free tier as "subject to change," and marks OVH AI Endpoints as Beta Access. Those hedges are honest, but they also mean the table is a snapshot of provider policy that only the provider can revoke.

The failure mode is predictable: you build against a row, the provider changes its free tier, and your code fails at the API call rather than at review time. Nothing in the repository can detect that for you, because the repository does not proxy requests or check quotas. It is documentation plus snippets.

A second limitation is that rate limits in the table are not comparable across rows. Google AI Studio is listed as 5-30 RPM depending on model, with 9,000 requests per day for Flash but 25 per day for Gemini 3.1 Pro. Mistral is listed as 1 request per second with no daily figure. Hugging Face Inference is capped by a monthly credit of $0.10 rather than a request count, with 300 requests per hour. Reading those as equivalent numbers would be a mistake; the README presents them as the providers publish them.

It is also the wrong tool if you need a model that is not on the list, or if your workload needs throughput beyond what any free tier allows. The README makes no claim about paid tiers or self-hosting beyond mentioning local tools in its header.

Free-LLM compared with a single-provider free tier

The obvious alternative is to skip the directory and sign up directly with one provider that has a permanent free tier, such as Groq or Google AI Studio. The difference is scope and maintenance burden. A single provider gives you one set of terms to track, one key to rotate, and one support channel, at the cost of being stuck if that tier changes. Free-LLM gives you a comparison surface across 41 providers and code examples for each, at the cost of having no control over any of them.

A second alternative is running a model locally. The README's own header mentions "9 local/self-hosted tools for unlimited private use," which is the opposite trade: no rate limits and no third-party data handling, but you supply the hardware and accept whatever quality fits on it. Free-LLM lists those tools rather than running them, so it does not help you pick hardware.

The comparison page at free-llm.com/compare is the part of the project that most resembles a decision tool, since it puts two providers side by side on limits, models and pricing. That is a web feature, not something the repository provides offline.

Licence, maintenance and the cost of keeping up

The repository is MIT licensed, with a LICENSE file at the top level. That covers the repository's own content and code examples. It does not cover the providers, the models, or the APIs the directory points at, each of which carries its own terms. Copying a snippet from code-examples/ into your project is covered by MIT; calling the endpoint the snippet targets is governed by that provider's agreement, and some free tiers restrict commercial use or require attribution. Check the provider, not this repository.

The last push to the repository was on 2026-09-17, and the repository is not archived. The README states the content is synced daily from free-llm.com, so the practical upgrade path is to pull the repository or read the site rather than to track a version number. There are no releases retrieved, which fits a documentation-first project: there is nothing to pin.

Translation maintenance is the hidden cost. The repository ships README files in English, Simplified Chinese, Traditional Chinese, Japanese and Korean. Every change to the provider tables has to propagate to five files, and the count discrepancy between "120+ models from 41 providers" and "all 33 providers" suggests that propagation is not perfectly synchronized.

Editorial conclusion

Adopt Free-LLM if you need a shortlist of free LLM API tiers and a working OpenAI-compatible snippet to start from; the repository states it is synced daily from free-llm.com, so the directory is the part worth trusting and the provider terms are the part worth re-checking. Skip it if you need a guaranteed quota for production traffic, or if you want a single vendor with a support contract. Before you build anything on top of a provider row, verify two things at the provider itself: whether the free tier still exists on the date you deploy, and whether the licence and data-handling terms of the model you picked allow your use case. The repository's own tables list card requirements and limits as of the last sync, not as a commitment.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. nejib1/Free-LLM on GitHub
4. Project website
5. README