mnfst/awesome-free-llm-apis

List of Permanent Free LLM API (API Keys)

This is a curated list (the "awesome" format) of APIs for running large language models (LLMs) that have permanent free tiers. It is a reference guide — it does not host any models itself; it links to each provider so you can sign up and get an API key. The blurb also mentions that these same free APIs are available through a related product called manifest.build (a tool for "building reliable agents"), but the README does not explain further what that product does beyond that single sentence.

The list is split into two groups, and every entry includes the same key facts:

  • Provider APIs — APIs run by the companies that trained or fine-tuned the models themselves. Each entry lists the company (e.g., Aion Labs, Cohere, Google Gemini, Mistral AI, Z AI), its country, the base URL to call, a table of model names with their context length, max output token count, modality (text-only vs. text + image, etc.), and rate limits.
  • Inference providers — third-party platforms that host open-weight models from various sources. Examples are Cloudflare Workers AI, Groq, Hugging Face, Kilo Code, LLM7.io, ModelScope, NVIDIA NIM, Ollama Cloud, OpenRouter, OVHcloud, and more. They follow the same format: provider, country, base URL, model tables, rate limits.

Each entry also states the free-tier conditions clearly, and those conditions vary widely. For example: Google's free tier requires no credit card but may use your prompts to improve their products; Mistral's free mode gives $10/month in credits but may use prompts for training unless you opt out; Cohere's trial is non-commercial use only with 1,000 calls/month; Cloudflare gives 10,000 "neurons" per day; Groq offers a free tier without a card; Hugging Face gives $0.10/month in credits (subject to change); ModelScope requires real-name verification; OVHcloud has a free anonymous tier with no signup at all. Some entries note specific limits (e.g., 2 RPM per IP for anonymous OVH access) or that certain models are "retirement announced" (GLM-4.5-Flash).

The README makes one global note: all endpoints are OpenAI SDK-compatible unless stated otherwise, so code written for OpenAI's API should work with these providers with just a URL and key change.

Finally, the list has a short glossary section mentioned in its table of contents (Contents), though the README text included here does not show the glossary contents.

Related

  • Project
    triton-inference-server/server

    An open-source inference serving software that streamlines the deployment of AI models from multiple frameworks across cloud, data center, edge, and embedded devices.

  • Project
    openvinotoolkit/openvino

    An open-source toolkit for optimizing and deploying deep learning models across CPU, GPU, and NPU hardware to boost inference performance.

  • Project
    xlite-dev/LeetCUDA

    LeetCUDA is an open-source book and repository of 200+ high-performance CUDA kernels designed to teach developers how to implement optimized GPU operations like FlashAttention and GEMM.

  • Project
    LMCache/LMCache

    LMCache is a vendor-neutral KV cache management layer that offloads and persists LLM inference cache to improve throughput and reduce latency for long-context and agentic workloads.