# Chat completions

> Generate text with DiffusionGemma 26B. The request and response follow OpenAI's chat completions format.

Source: https://codiv.ai/docs/api-reference/chat-completions

## Create a chat completion

Endpoint:


```text
POST https://api.codiv.ai/v1/chat/completions
```

Authenticate with `Authorization: Bearer <key>`. Usage counts toward your [text-generation quota](https://codiv.ai/docs/guides/rate-limits.md).

### Request body

- `model` (string, required)

  `diffusiongemma-26b`.

- `messages` (array, required)

  The conversation, as OpenAI message objects (`system`, `user`, `assistant`, `tool`). User content can include `image_url` parts.

- `max_tokens` (integer, optional)

  Most tokens to generate. Defaults to 1024, capped at 8192. `max_completion_tokens` is accepted too.

- `response_format` (object, optional)

  `{"type": "json_object"}` or `{"type": "json_schema", "json_schema": {...}}`. The model is told to reply with JSON only, and the first JSON object in its reply is returned as `content`. The schema guides the model but is not enforced.

- `stream` (boolean, optional)

  Stream server-sent events. The last event before `[DONE]` carries `usage`.

- `tools` (array, optional)

  Function tools, with `tool_choice` `"auto"`, `"none"` or a named function.

- `chat_template_kwargs` (object, optional)

  `{"enable_thinking": true}` lets the model think before answering. Off by default.

Ignored: `temperature`, `seed`, `min_p`, `logit_bias`, `presence_penalty`, `frequency_penalty` and `reasoning`. A diffusion model has no per-token sampling for them to control.

### Returns

A chat completion object. `message.content` is the reply; `message.reasoning` holds the thought when thinking is on. `usage.prompt_tokens` and `usage.completion_tokens` are what count toward your quota.

curl:

```bash
curl https://api.codiv.ai/v1/chat/completions \
  -H "Authorization: Bearer $CODIV_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "diffusiongemma-26b",
    "messages": [{"role": "user",
      "content": "Name three fruits as JSON: {\"fruits\": [...]}"}],
    "response_format": {"type": "json_object"},
    "max_tokens": 256
  }'
```

Python:

```python
from openai import OpenAI

client = OpenAI(base_url="https://api.codiv.ai/v1",
                api_key="sk-codiv-...")
r = client.chat.completions.create(
    model="diffusiongemma-26b",
    messages=[{"role": "user",
      "content": "Name three fruits as JSON: {\"fruits\": [...]}"}],
    response_format={"type": "json_object"},
    max_tokens=256,
)
print(r.choices[0].message.content)
```

Response:


```json
{
  "id": "chatcmpl-9f2c41",
  "object": "chat.completion",
  "model": "diffusiongemma-26b",
  "choices": [{
    "index": 0,
    "message": {"role": "assistant",
      "content": "{\"fruits\": [\"apple\", \"mango\", \"kiwi\"]}"},
    "finish_reason": "stop"
  }],
  "usage": {"prompt_tokens": 41, "completion_tokens": 64,
            "total_tokens": 105}
}
```

## Errors
Errors use OpenAI's shape:

Error:


```json
{"error": {"message": "...", "type": "rate_limit_error", "code": "rate_limit_exceeded"}}
```

| Status | type | When |
|---|---|---|
| 400 | `invalid_request_error` | The request is malformed, or the model rejected it |
| 401 | `authentication_error` | Unknown or revoked key |
| 403 | `authentication_error` | No key sent |
| 403 | `permission_error` | The account is disabled or its email is not verified |
| 413 | `invalid_request_error` | Body larger than 8 MB |
| 429 | `rate_limit_error` | Too many requests per minute on one key |
| 429 | `insufficient_quota` | The text-generation quota is used up |
| 503 | `api_error` | The model failed on this request |
| 529 | `overloaded_error` | All generation slots are busy. Retry after `retry-after`. |

Generation denoises the reply in 64-token blocks, so a completion takes seconds, not milliseconds. Retry `429 rate_limit_error` and `529` with backoff; don't retry `insufficient_quota`.