We aggregated 190+ free AI models into one endpoint

UnoRouter aggregates free lanes from many upstream providers into more than 190 models with a genuine free tier behind one key. Each free model allows about 1 request per minute per user; hitting the cap returns HTTP 429 with a Retry-After header, and drained pools return an explicit all providers busy error and recover automatically. Paid models bill pay as you go from $1.

The free-LLM landscape is real but scattered: Groq, Gemini, Cloudflare, NVIDIA, SiliconFlow and a dozen others each give away genuine capacity, behind a dozen signup pages, a dozen key formats, and a dozen incompatible APIs. We discovered, tested, and merged every legitimate permanent-free provider we could into UnoRouter. The result: 190+ free model rows from 18 providers behind one OpenAI-compatible endpoint and one key.

What we added

Eighteen free providers, one at a time: Groq, Gemini, Cerebras, SambaNova, Mistral, Cloudflare Workers AI (two accounts), GitHub Models, Z.ai, OVHcloud, AI Horde, Pollinations, Cohere, Jina, NVIDIA NIM (full open catalog: Nemotron, Llama, Qwen, gpt-oss, DeepSeek and more), SiliconFlow (DeepSeek-V4, GLM-5.1, Qwen3.5/3.6, Kimi-K2.6, MiniMax-M3), the HuggingFace router, and LLM7. That is 190+ free model rows: Llama, gpt-oss, Qwen, Mistral, GLM, Nemotron, DeepSeek and more, plus free embedding, image, and audio models. Every one is probed end to end for HTTP, streaming, and tool calls before it goes live, the same authenticity and harness checks we run on paid models.

Free for a reason

These models are genuinely free, and that is exactly why they have limits. Each upstream enforces its own: requests per minute, daily token quotas, Cloudflare neuron budgets, volunteer-queue priority. Hit a cap and that provider returns 429 until it resets. A free key that worked this morning can be exhausted by this afternoon. Free tier is best-effort throughput, not a guarantee. If your workload needs predictable latency and no surprise 429s, use a paid model.

Our own rate limit, and why it is there

On top of the upstream caps we add a small limit of our own: one request per minute per free model, per user. It is deliberate. A single free model is often one shared upstream pool of maybe a million tokens a day for everyone combined, so without a cap a handful of heavy users would drain it in minutes and everyone else would see nothing. One per minute per model keeps the pool alive and spread across the whole community, and you can spread traffic across the many free models rather than hammering one. When you hit our cap you get an HTTP 429 with a Retry-After header telling you exactly how long to wait. This is separate from the upstream 429 and from the 503 below. If you want higher throughput on a specific model, use its paid tier where the limit does not apply.

Why aggregate them at all

Because the alternative is eighteen accounts. Each provider has its own signup, its own key format, its own base URL, and its own quirks: Z.ai speaks the Zhipu V4 path, Cloudflare carries the account id in the URL, AI Horde wants an anonymous key, GitHub gates models behind a token scope, the HuggingFace router fans out across a dozen upstreams. We absorbed all of that so you call them the way you call everything else: one OpenAI-compatible endpoint, one key, a model name. The honest rule we hold ourselves to: one real account per provider, caps accepted, nothing farmed, nothing pooled. We expose the free tier as a gift, not as a resale of someone else's quota.

How we soften the limits

Many of these models are served by more than one free provider. Llama 3.3 70B alone runs on seven of them. When several providers offer the same model, we collapse them into one published name and fail over automatically: if one upstream returns a 429 or goes quiet, the next healthy provider serving that model takes the request. That is the one lever we do control. If every provider for a model is momentarily rate-limited, the request returns an HTTP 503 with an explicit "all providers busy" message (not a 404, and not "model not found"), and the channel is re-tested and re-enabled within minutes by our health cron. So a multi-source free model keeps answering long after any one of its providers is exhausted. Single-source free models have no fallback, so they stall when their one upstream is spent.

What we did not do

We did not add reverse proxies that re-serve OpenAI or Claude flagships without permission. We did not pull in personal-key aggregators whose tokens are non-transferable, or pool-of-pools services that farm and rotate other people's keys. Those exist and they are tempting and they are exactly the gray-market mess this gateway is meant to replace. Every provider on the list gives its free tier away on purpose, under its own terms. If a source could not pass that bar, it is not here.

Try it

All 190+ free models are live behind one OpenAI-compatible endpoint. Grab an API key or browse the model catalog and filter for free. Just remember which tier you are on when a 429 or 503 shows up.