Familiar story: you got excited about the idea of building a bot or smart document search, opened the console of a major cloud provider, linked your card, and a week later found charges for test runs. Worst case, a script left running in a loop burns through the monthly limit in a couple hours.

Yet for most prototypes, local experiments, and home servers, you don't need to pay at all. A repository by developer 12britz has been gaining popularity on GitHub. It's a guide to open weights, free APIs that don't require card details, and software for local inference.

What makes this collection different from dozens of others

"Awesome" lists on GitHub often die within six months of creation. Links go stale, free tiers shut down, and authors abandon maintenance. The creator of went a different route: they set up regular checks of live links and remove dead projects.

The database contains around 50 open models, over 200 tools, and more than 20 sections: from ready-made interfaces and routers to embeddings, vector databases, and voice engines. The main criterion is simple: using it shouldn't require money.

What's useful inside

The entire repository is broken down into practical categories. Filtering for the most interesting parts gives you a ready-to-use set for building applications at any level.

Free cloud APIs

If you have a weak laptop or don't want to deal with GPUs, you can call ready-made endpoints:

• Google AI Studio offers the most generous free tier on the Gemini Flash model lineup, which is more than enough for prompt debugging and testing.
• Groq delivers open model inference at massive speed thanks to custom LPU chips, with the base tier requiring no payment.
• NVIDIA NIM provides access to accelerated versions of popular open models directly on NVIDIA server infrastructure.
• OpenRouter and AnyAPI give a unified interface where you can filter for free models only and switch between them through a single config.

Local run on your own hardware

For those who keep data close or build software for closed environments, the list includes open weights and inference engines:

• Engines like Ollama, LM Studio, and llama.cpp install in a couple clicks and handle quantization and memory distribution.
• For servers and high loads, there's vLLM and SGLang with efficient context distribution support.
• Open weights include smaller options for regular laptops like Gemma, Phi-4, or Qwen, as well as heavy MoE architectures on the level of DeepSeek or Mistral.

Code assistants and terminal agents

A section for those looking to replace paid Copilot or Cursor subscriptions:

• Aider works directly in the terminal and automatically commits changes to a local Git repository.
• Cline integrates into VS Code, can run commands in the console, and create files step by step.
• OpenHands and Goose handle complex engineering tasks autonomously.

Routers and request optimizers

If you're connecting multiple free tiers at once, you need to manage them somehow. Projects like LiteLLM, 9Router, and OmniRoute handle traffic proxying, prompt compression, and failover to a backup provider if the primary one hits the daily limit.

Pitfalls of free tiers

Free cheese comes with its own limitations, which the author honestly warns about in the source descriptions.

The first issue involves requests per minute (RPM) and per day (RPD) limits. On free tiers, they're strict. If your agent starts actively calling tools in a loop, you'll quickly hit a error.

The second detail relates to data confidentiality. Some cloud providers under the free tier reserve the right to use incoming prompts for training future models. If you're working with proprietary company code or sensitive customer data, you can't use free cloud keys—it's safer to run quantized models locally through Ollama or vLLM.

Finally, model availability in free clouds often fluctuates. Today a service offers a model for free, tomorrow it restricts access due to server overload. For stable operation, configure a router with fallback to 2-3 alternative gateways from the start.

How to apply in practice

The simplest scenario: set up a combination of Aider in the terminal, the LiteLLM proxy router, and a free key from Google AI Studio or Groq. You get a full-featured code assistant within your workflow without a monthly fee.

If you need a RAG pipeline for knowledge base search, you can take the embedded database Chroma or LanceDB from the same list, hook up the open BGE-M3 embedding model, and spin up the Open WebUI web interface in a Docker container.

The repository will be useful for students, solo developers, and anyone who wants to build side projects without watching their account balance.

Bookmark the repository: https://github.com/12britz/awesome-free-models. Next time you need to spin up a model for testing, start with this list instead of entering credit card details into another cloud console.