DEV Community

albe_sf
albe_sf

Posted on

Mistral just shipped a vision model and cut flagship prices

#ai #machinelearning #opensource #llm

Mistral just made two significant moves that position it as an increasingly practical alternative to the major closed-source labs. The company released Pixtral, a new vision model, and simultaneously announced a price reduction for its flagship model, Mistral Large 2. This isn't just a routine update; it's a clear signal about their strategy: compete directly on multimodal features while aggressively pushing down cost.

what just shipped

The update, which appeared on their official changelog on September 17, included three key changes.

First, the release of pixtral-12b-2409, their new vision model. The naming suggests a 12-billion parameter model, a deliberate choice to offer strong performance in a cost-effective, easily deployable size. This follows a broader industry trend toward multimodality, but Mistral's commitment to open and efficient models makes this release particularly notable for builders.

Second, they released an updated version of their small model, mistral-small-2409. While less flashy than a new vision model, continuous improvement of smaller, efficient models is critical for production use cases where latency and cost are primary constraints.

Third, Mistral cut the price on their most capable model, Mistral Large 2, and introduced a free API tier on their platform, La Plateforme. This directly addresses one of the biggest barriers to adoption for smaller teams and individual developers: cost. Lowering the price of your top-tier model is a confident move designed to capture more production workloads.

competing on cost and capability

The release of Pixtral is a direct answer to the multimodal models from larger labs. For developers building applications that need to understand or process images, this provides a new, potentially more open and efficient option. While benchmarks are not yet available, a 12B parameter model is large enough for serious tasks without incurring the inference costs of massive frontier models.

Integrating a model like this into a workflow is straightforward. You can expect to interact with it through the standard API, likely with a modified payload to handle image inputs, such as a base64-encoded string or a URL.

import requests
import base64

API_KEY = "YOUR_MISTRAL_API_KEY"
MODEL_NAME = "pixtral-12b-2409"

# Encode a local image file
with open("path/to/your/image.jpg", "rb") as image_file:
    encoded_string = base64.b64encode(image_file.read()).decode('utf-8')

headers = {
    "Authorization": f"Bearer {API_KEY}",
    "Content-Type": "application/json",
}

data = {
    "model": MODEL_NAME,
    "messages": [
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "Describe the contents of this image in detail."},
                {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{encoded_string}"}}
            ]
        }
    ]
}

response = requests.post("https://api.mistral.ai/v1/chat/completions", headers=headers, json=data)

print(response.json())
Enter fullscreen mode Exit fullscreen mode

This is more than just adding a feature. It's about providing the core building blocks that engineers need. The simultaneous price cut on Mistral Large 2 reinforces this. It makes the entire stack, from small and efficient to large and powerful, more economically viable. For teams running systems at scale, these cost differences add up quickly.

the bigger picture: open distribution

Mistral's strategy appears to extend beyond just models and APIs. Their recent partnership with Mozilla to integrate Mistral models into the Firefox browser is a move to control distribution and reach users outside of the typical developer-focused cloud platforms. By building a presence directly in the browser, they are creating a new channel for their technology, one that is built on a foundation of open technology.

This matters. As AI becomes more deeply embedded in our daily tools, the question of who controls the underlying models becomes critical. By partnering with organizations like Mozilla and continuing to release open-weight models, Mistral is providing a real alternative to the closed ecosystems of Big Tech.

For builders, this is a positive development. It means more choice, better pricing, and the ability to build on platforms that align with the open principles of the web. The latest releases are not just new tools, but a continued investment in an ecosystem that offers more than one way to build.

Sources

Top comments (0)