awesome-free-llm-apis review: free LLM APIs worth verifying

In brief: mnfst/awesome-free-llm-apis is a curated index of LLM API providers that publish permanent free tiers, with endpoints, model names, context windows, and rate limits for each entry. At the 2026-08-21 review snapshot, the repository showed 6,852 stars and 658 forks, is licensed CC0-1.0, lists JavaScript as its primary language, and has no tagged releases. This is a source and documentation review (verification level: observed); use the list for discovery and prototyping, and re-check every limit and policy at the provider before production.

Fact	Value
Review date	2026-08-21
Repository	mnfst/awesome-free-llm-apis
License	CC0-1.0 (Creative Commons Zero v1.0 Universal)
Stars (snapshot)	6,852 stars, 658 forks at 2026-08-21
Primary language	JavaScript (per GitHub metadata)
Latest release	None found; no tagged releases in repository metadata
Verification level	Observed (source and documentation review)

What this project is

The README describes the project as “LLM APIs with permanent free tiers for text inference” and states that all endpoints are OpenAI SDK-compatible unless noted, with each link pointing to the provider’s API key page. It is an awesome list, not a framework: there is no runtime to install, deploy, or upgrade. The JavaScript language tag describes repository metadata, not a component your application would use.

The list is maintained under the mnfst organization, and the README carries a Manifest banner linking to manifest.build, which offers these same free LLM APIs. That sponsorship is relevant context when evaluating the list’s neutrality.

What the list covers

At the snapshot the README documents five provider APIs (Aion Labs, Cohere, Google Gemini, Mistral AI, Z AI) and eleven inference providers (Cloudflare Workers AI, Groq, Hugging Face, Kilo Code, LLM7.io, ModelScope, NVIDIA NIM, Ollama Cloud, OpenRouter, OVHcloud AI Endpoints, SiliconFlow). Provider APIs are run by the companies that train or fine-tune the models; inference providers host open-weight models from multiple sources.

Representative entries from the README, current as of the snapshot:

Google Gemini free tier: Flash-class models with up to 1M context at 15 to 30 RPM and 1,500 requests per day, with the caveat that Google no longer publishes per-model free-tier rate limits.
Mistral AI free mode: about $10/month in API credits, shared across Studio, the API, and Vibe Code; free-mode prompts may be used to train Mistral models unless you opt out.
Cohere free trial: 1,000 API calls per month, non-commercial use only.
OpenRouter: 17 free models (marked :free) at 20 RPM and 50 requests per day each; a one-time $10 credit purchase unlocks 1,000 requests per day.
Groq: free tier at roughly 30 RPM and 1,000 requests per day for most listed models, with compound models at 250 per day.
Cloudflare Workers AI: 10,000 neurons per day shared across all Workers AI usage, with overage failing the request rather than billing.
OVHcloud AI Endpoints: anonymous tier at 2 requests per minute per IP per model, no signup or API key.

Who should use it

The list fits a founder, developer, or operations person prototyping an AI agent, chatbot, summarization, classification, or small internal automation who wants a low-cost pool of providers without signing up for every vendor individually. It also works as a comparison sheet for context windows, output limits, and rate ceilings before committing to a paid plan.

It is not a production contract. If your workload needs a stable SLA, consistent throughput, confidentiality, or commercial guarantees, expect to move to a paid tier or a direct agreement, and verify the provider terms yourself.

What to check before production

Every row in this list is a provider claim with a different verification date. The README itself flags several moving parts as of August 2026:

Retirements and catalog rotation. Groq shut down llama-3.3-70b-versatile and llama-3.1-8b-instant on 2026-08-16; LLM7.io rotated its catalog in August 2026 so previously listed models return model_unavailable; Kilo Code’s free pool changes frequently and its catalog can lag what is actually served; Z AI announced retirement of GLM-4.5-Flash while it remains catalogued. Check the provider page before wiring a model in.
Data use. Google’s free-tier prompts may be used to improve products. Mistral’s free mode may use inputs and outputs for training unless you opt out. OpenRouter notes free providers may log prompts for training. Kilo Code’s docs warn its auto-router “may route your requests to providers that log prompts and outputs”. For customer or confidential data, these terms change the whole provider choice.
Commercial restrictions. Cohere’s free trial is non-commercial only; other rows may carry their own use restrictions.
Account requirements. ModelScope requires Alibaba Cloud account binding and real-name verification; SiliconFlow requires real-name verification and international users must contact support; NVIDIA NIM requires Developer Program membership; Z AI’s chat API does not require real-name verification while its Batch API does.
Quota behavior. Confirm whether exceeding a limit fails the request, pauses it, or starts billing. Cloudflare’s free neurons fail over; OVH’s anonymous tier is capped per IP per model with pay-as-you-go tiers above it.

Maintenance signals

GitHub metadata at the snapshot shows the repository was created on 2026-03-21, last pushed on 2026-08-20, with 6,852 stars, 658 forks, and 6 open issues. The README carries dated footnotes (Groq deprecations as of 2026-08-16, Kilo Code rows verified with live requests on 2026-08-19, LLM7 catalog rotation in August 2026), which indicates active curation rather than a stale list. The main maintenance risk is structural: there are no tagged releases, so there is no versioned snapshot, and the value of the list depends on the maintainers and contributors keeping hundreds of provider rows current.

Deployment and integration concerns

Since the project is a directory, deployment means connecting your application to one or more listed endpoints. The OpenAI SDK-compatible base URLs simplify swapping providers at the client level, but the endpoints differ in shape: Google’s generativelanguage.googleapis.com/v1beta, Cohere’s api.cohere.com/v2, SiliconFlow’s api.siliconflow.cn/v1, and Cloudflare’s account-scoped path, so a small adapter or config layer is still part of the work. Rate limits are per provider and often per model, so 429 handling, retries, backoff, key storage, and fallbacks are on your side. The README points to routing options such as OpenRouter’s free-models router and model fallbacks and Kilo Code’s auto-router, and the repository topics include llm-router and llm-routing.

Limitations and evaluation questions

Do not mistake coverage for proof. This review includes no benchmarks, no uptime data, and no runtime evidence; rows vary in how recently they were verified, and the list policy excludes trial credits and time-limited promos, so it understates the full menu of cheap options. The commercial link to manifest.build should be weighed when judging curation.

Before building on any entry, ask: what are the current per-model limits and are they published or estimated? Is commercial use allowed and is data used for training? What identity or account verification is needed for your region? What happens at quota exhaustion? Is there a paid tier with the same API contract when usage grows? And who last verified this row: the maintainer, a contributor, or the provider?

Verdict

mnfst/awesome-free-llm-apis is a genuinely useful discovery catalog for free LLM APIs, and the maintenance pattern at the snapshot is healthy. For low-volume agent prototypes and quick provider comparisons, it can save real time. For production, treat it as a pointer: verify limits, data policies, and retirement notices on the provider’s own pages, and design routing and fallbacks so no free tier becomes a single point of failure.

What I can do next

The work after this list is exactly where the repository stops: matching providers to your traffic and data-use constraints, wiring the OpenAI-compatible endpoints into your application, adding routing, retries, fallbacks, and key management, and monitoring quotas before they bite. That is the AI and automation work I do at XP812 with small teams, including taking over an existing free-tier integration that needs a more stable path.