Pixtral 12B

Deploy Pixtral 12B multimodal model on Koyeb’s high-performance cloud infrastructure.

With one click, get a dedicated GPU-powered inference endpoint ready to handle requests with built-in autoscaling and scale-to-zero.

Overview of Pixtral 12B

Pixtral 12B is a multimodal model with 12 billion parameters, complemented by a 400 million parameters vision encoder. The model delivers strong capabilities in tasks such as chart and figure comprehension, document question answering, multimodal reasoning, and instruction following.

Pixtral 12B will be served with the vLLM inference engine, optimized for high-throughput and low-latency model serving.

The default GPU for running this model is the Nvidia A100 instance type. You are free to adjust the GPU instance type to fit your workload requirements.

Quickstart

Pixtral 12B one-click model is served using the vLLM engine. vLLM is an advanced inference engine designed for high-throughput and low-latency model serving. Optimized for large language models, it provides efficient performance and compatibility with the OpenAI API.

After you deploy the Pixtral 12B model, copy the Koyeb App public URL similar to https://<YOUR_DOMAIN_PREFIX>.koyeb.app and create a simple Python file with the following content to start interacting with the model.

import os

from openai import OpenAI

client = OpenAI(
  api_key=os.environ.get("OPENAI_API_KEY", "fake"),
  base_url="http://localhost:8000/v1",
)

chat_completion = client.chat.completions.create(
  messages=[
    {
      "role": "user",
      "content": [
        {
          "type": "image_url",
          "image_url": {
            "url": "https://images.unsplash.com/photo-1506744038136-46273834b3fb"
          },
        },
        {"type": "text", "text": "Describe the image."},
      ],
    },
  ],
  model="mistralai/Pixtral-12B-2409",
  max_tokens=50,
)

print(chat_completion.to_json(indent=4))

The snippet above is using the OpenAI SDK to interact with the Mistral 7B model thanks to vLLM OpenAI compatibility.

Take care to replace the base_url value in the snippet with your Koyeb App public URL.

Executing the Python script will return the model's response to the input message.

python main.py

{
    "id": "chatcmpl-953c053e41c346258c2e654eb228024c",
    "choices": [
        {
            "finish_reason": "length",
            "index": 0,
            "logprobs": null,
            "message": {
                "content": "The image captures a tranquil scene of a mountainous landscape at sunset. The sky, awash with hues of pink and orange, casts a warm glow over the scene. The mountains, shrouded in a veil of fog, add an air of mystery to",
                "role": "assistant",
                "tool_calls": []
            },
            "stop_reason": null
        }
    ],
    "created": 1737132182,
    "model": "mistralai/Pixtral-12B-2409",
    "object": "chat.completion",
    "usage": {
        "completion_tokens": 50,
        "prompt_tokens": 2803,
        "total_tokens": 2853,
        "prompt_tokens_details": null
    },
    "prompt_logprobs": null
}

Securing the Inference Endpoint

To ensure that only authenticated requests are processed, we recommend setting up an API key to secure your inference endpoint. Follow these steps to configure the API key:

1. Generate a strong unique API key to use for authentication
2. Navigate to your Koyeb Service settings
3. Add a new environment variable named VLLM_API_KEY and set its value to your secret API key
4. Save the changes and redeploy to update the service

Once the service is updated, all requests to the inference endpoint will require the API key.

When making requests, ensure the API key is included in the headers. If you are using the OpenAI SDK, you can provide the API key through the api_key parameter when instantiating the OpenAI client. Alternatively, you can set the API key using the OPENAI_API_KEY environment variable. For example:

OPENAI_API_KEY=<YOUR_API_KEY> python main.py