FreeLLMAPI: 34 free big model manufacturers twisted into one OpenAI interface, about 7.4 billion tokens per month freeFreeLLMAPI：把 34 家免费大模型厂商拧成一个 OpenAI 接口，每月约 74 亿 token 免费用

The pain point of the free quota is "scattered": Google, Groq, Cerebras, Mistral, OpenRouter, Cloudflare, Cohere, Z.ai (Smart Spectrum), NVIDIA, HuggingFace, but to connect at the same time in your code, the workload is much larger than the paid API.

And the free ecosystem is changing every week: manufacturers launch new models, offline old models, and quietly change quotas. The router of FreeLLMAPI will pull the signed model directory by itself, and it can keep up with changes without git pull, eliminating long-term maintenance.

After the aggregation, there is an additional benefit: the single house will automatically retry the next house (with cooling and key rotation), which is equivalent to turning "34 unstable small faucets" into a stable water pipe.

The most practical meaning for individual developers is that it used to take several accounts and several adaptation layers to get the computing power, but now it only needs to point to one localhost endpoint.

Built-in free vendors include Google, Groq, Cerebras, OpenCode Zen, Mistral, OpenRouter, Cloudflare, Cohere, Z.ai (Smart Spectrum), NVIDIA, HuggingFace, and ModelScope (Qwen3/DeepSeek V4/GLM-5) that requires Alibaba Cloud binding, and more than 22 free vendors not listed in the main table.

A custom vendor is also supported: point the chat/embedding/image/audio model to any OpenAI-compatible endpoint, including llama.cpp, LM Studio, vLLM, or local Ollama.

The client side covers mainstream coding agents: Claude Code, Codex CLI, Gemini CLI, Cursor, Cline, Roo Code, OpenCode, Aider, and Continue, Goose, Qwen Code, Kilo Code, Crush, Zed, JetBrains AI, OpenClaw, Hermes Agent, etc., as long as they support OpenAI compatible, Anthropic, Gemini or Ollama protocols.

Most agents can complete the configuration with one command - it will read your current directory, backup the existing configuration and then merge it into it, leaving your original settings unchanged.

Full coverage of OpenAI-style interfaces: chat/completions, responses (required by Codex CLI), completions (editor gray completion), images/generations, videos/generations, audio/speech, audio/transcriptions, embeddings, models, all support streaming and non-streaming.

Anthropic Messages API:/v1/messages speaks directly to Anthropic's online format, so Claude Code and the official Anthropic SDK can run directly on your heap of free credits.

Native Gemini and Ollama interface: The Gemini CLI can use/v1beta, and the optional Ollama simulation layer can match local model clients such as Zed and JetBrains.

Fusion multi-model synthesis: Request a virtual fusion model, the router will send the question to a set of complementary free models in parallel, and then a referee model will synthesize an answer.

Six routing strategies: Sort the model by real-time speed/capability/stability, and automatically downgrade to the next one with cooling and key rotation at 429 or 5xx.

The tool call and structured output are transparently passed back and forth in OpenAI format (the tool call in plain text format will be recovered as a real tool_calls), and the response_format, seed, logprobs and other sampling parameters are transparently passed one by one according to the manufacturer's ability.

The following is the complete source code package of the server-side repository. It is recommended to use Docker self-hosting. The repository comes with compose configuration and installation documents.

After starting the local service (default port 3001), first fill in the API Key of each free vendor on the Keys page, and the router will automatically pull the available model directory.

Then configure your client with a unified local address and key, for example: npx freellmapi setup-claude --url http://localhost: 3001 --api-key unified key, which automatically registers the configuration of Claude Code.

In addition to the source code, the official macOS/Windows desktop, Docker image, Google Play and App Store apps are also available; the free version gets a monthly model snapshot (a model enters the free version 30 days after being added to the real-time directory), and the paid version ($19/year) is synchronized on the same day.

Note: The difference between the free version and the paid version is only in the update timeliness of the model directory, and the routing and aggregation capabilities are consistent; there is also a disclaimer on the terms of use of each manufacturer in the document, which is recommended to be read first and then used on a large scale.

This site organizes resources from GitHub open source repository tashfeenahmed/freellmapi (mit license), for learning and technical exchanges only. When aggregating free credits, please abide by the terms of service and rate limits of each manufacturer, and do not use the free credits for commercial resale. Although the keys are encrypted and stored locally, it is still recommended to use a separate, revocable Key at any time. In this article, "about 7.4 billion tokens per month" is the self-explanatory caliber of the project README, and the actual available amount changes with the manufacturer's policy, subject to the latest official instructions.