Llama API relay services 16 in total

Llama has the most complete ecosystem of any open-weight family, and its presence on relays comes mainly from one need: calling open models such as Llama, Mistral and Qwen through the same interface as closed-source flagships.

The thing to confirm first with a Llama relay is the model version and quantisation tier. One generation of an open model comes in several parameter sizes (8B, 70B, 405B and so on) and different quantisation levels, whose cost and quality differ substantially — while some services list only "Llama" with no specification. Pulling the real ids from /v1/models is an essential check.

Another practical consideration: if all you need is an open model, hosting it yourself is often the better deal. The point of an open licence is that you can hold it — once volume is high enough, renting inference costs more than running your own. A relay suits the stage where usage is small and you want to validate quickly.

Our Open source / self-hosted category lists self-hosting options such as One API, New API and LiteLLM, which you can compare against the relays here to see which fits your stage.

• Most stable：Highest current uptime. Where samples are thin we show the public-dataset baseline instead, and the detail page states the source.
• Fastest：Home-page response time from the most recent probe. A single measurement is noisy, so treat it as a range rather than an absolute.
• Broadest model coverage：This uses the model count our endpoint probe actually detected, not the support list the service advertises — the two are measured differently, and the advertised figure is usually the more optimistic one.

Should I use a relay for Llama or deploy it myself?

It depends on the stage. Small usage and a quick validation — use a relay, the barrier is lowest. Large, steady usage — self-hosting works out better, because the cost of holding an open model is fixed while pay-as-you-go grows linearly with volume. Our Open source / self-hosted category compares the self-hosting tools.

How do I confirm which specification of Llama a service is providing?

Request /v1/models and look at the returned ids — they usually carry a parameter size (70b, 8b and so on). If the list says only "Llama" with no specification, ask their support directly, because cost and quality differ substantially between sizes.

Do open models get downgraded on relays?

Downgrading an open model usually means swapping in a smaller parameter size or a lower quantisation level, which noticeably degrades quality. To check: run the same hard task across services and compare output, or verify the exact specification returned by /v1/models.

How do I choose between Llama and other open models?

By task type. Llama has the most mature ecosystem, tooling and documentation; Mistral is more efficient on some reasoning tasks; Qwen is usually better on Chinese-language work. If you are unsure, run a batch of real tasks on each and compare accuracy — do not go by benchmark scores alone.