> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tera.gw/llms.txt
> Use this file to discover all available pages before exploring further.

# OpenAI compatibility

> What carries over from OpenAI clients, and what's different.

Tera implements the OpenAI Chat Completions API surface. Existing OpenAI SDKs work by changing two settings:

```python theme={null}
from openai import OpenAI

client = OpenAI(
    base_url="https://api.tera.gw/v1",
    api_key="sk-tera-...",
)
```

## What's the same

* **Endpoints** — `/v1/chat/completions`, `/v1/completions`, `/v1/models`, `/v1/audio/speech`
* **Streaming** — Server-Sent Events with `data: {...}` frames terminated by `data: [DONE]`
* **Request shape** — `messages`, `temperature`, `top_p`, `max_tokens`, `stop`, `seed`, `frequency_penalty`, `presence_penalty`, `stream`, `tools`, `tool_choice`, `response_format`
* **Response shape** — `id`, `object`, `created`, `model`, `choices[].message`, `choices[].finish_reason`, `usage`
* **Tool calling** — OpenAI-compatible `tools` array and `tool_calls` in responses

## What's different

### Model IDs

Tera uses HuggingFace IDs as the canonical model name — no provider prefix.

```json theme={null}
{ "model": "Qwen/Qwen2.5-7B-Instruct" }
```

See [Models](/models/overview) for the catalog.

### Extra sampling parameters

Tera supports a few sampling parameters beyond OpenAI's surface. They're optional and ignored if you don't pass them.

* `top_k` — top-k sampling
* `repetition_penalty` — additional penalty term (distinct from OpenAI's frequency/presence penalties)
* `min_p` — minimum probability threshold

### Reasoning models

Models like `Qwen/Qwen3.5-27B` emit explicit reasoning traces. Tera splits these into a separate `reasoning_content` field rather than mixing them with the user-facing answer. See [Reasoning](/concepts/reasoning).

### No org / project headers

We don't require or accept `OpenAI-Organization` or `OpenAI-Project` headers. Drop them if your client sends them — they're ignored.

### No moderation endpoint

We don't offer `/v1/moderations`. Use OpenAI's moderation endpoint if you need it, or run a separate guard model.

### No embeddings, no images, no fine-tuning

Today: text generation and TTS only. No `/v1/embeddings`, no `/v1/images/*`, no `/v1/fine_tuning/*`. Let us know if you need these.

## Behavioral notes

* **First request after cold start is slower** (\~2–12s TTFT) due to CUDA graph compilation. Subsequent requests are fast.
* **5xx errors trigger automatic retry** within the gateway across healthy backend replicas before being returned to you.
* **Health-aware routing** — if a backend fails health checks, traffic is steered to healthy replicas with no client changes.