Pixtral 12B: Mistral's First Vision-Language Model

Mistral Enters the Vision Space

Pixtral 12B is Mistral AI's first vision-language model - a 12-billion parameter multimodal model that combines a new vision encoder (400M params) with the Mistral Nemo 12B language backbone.

The model is released under Apache 2.0 - fully open for commercial use, fine-tuning, and deployment without restrictions.

Arbitrary Resolution Image Processing

Most vision-language models process images at a fixed resolution (e.g., 336×336 or 448×448 pixels), which distorts or loses information in images with unusual aspect ratios. Pixtral uses a variable-resolution approach: images are tiled into patches that preserve the native aspect ratio, with no forced resizing.

This matters for:

• Documents - preserve text readability at native resolution
• Panoramic images - no distortion from forced square crop
• Diagrams - fine detail in technical drawings remains crisp
• Charts - axis labels and data points remain legible

Benchmark Results

Pixtral beats LLaVA-1.6 34B (nearly 3x larger) on every benchmark - a strong result for a 12B model.

Using the Mistral API

from mistralai import Mistral
import base64

client = Mistral(api_key="your-api-key")

# Load image as base64
with open("chart.png", "rb") as f:
    image_b64 = base64.b64encode(f.read()).decode()

response = client.chat.complete(
    model="pixtral-12b-2409",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "image_url",
                    "image_url": f"data:image/png;base64,{image_b64}"
                },
                {
                    "type": "text",
                    "text": "Extract all data values from this bar chart as a JSON array."
                }
            ]
        }
    ]
)
print(response.choices[0].message.content)

Self-Hosting With vLLM

pip install vllm

# Start vLLM server with Pixtral
python -m vllm.entrypoints.openai.api_server     --model mistralai/Pixtral-12B-2409     --tokenizer_mode mistral     --config_format mistral     --load_format mistral     --max-model-len 32768     --gpu-memory-utilization 0.9

The vLLM server exposes an OpenAI-compatible endpoint - switch base_url in your existing code.

Use Cases

Invoice and receipt processing: Extract line items, totals, and vendor details from scanned documents with high accuracy.

Technical diagram analysis: Parse architecture diagrams, circuit schematics, or flowcharts and convert to structured descriptions.

Chart data extraction: Pull numeric values from bar charts, line graphs, and tables for automated reporting pipelines.

Screenshot understanding: Analyze UI screenshots for accessibility audits, bug reports, or automated testing.

Summary

Pixtral 12B is the strongest Apache 2.0 licensed vision model in its size class. The variable resolution approach is a genuine architectural improvement over fixed-patch models. Access it via Mistral API or download weights from HuggingFace.