Mistral released its first multimodal AI model, the Pixtral 12B, on Wednesday. The AI company, known for its open-source Large Scale Language Models (LLM), has made its latest AI models available on GitHub and Hugging Face for users to download and test. Although Pixtral is multimodal, it can only process images and answer questions about them using computer vision technology. Two special encoders have been added for this function. It cannot produce images like stable diffusion models or midjourney generative adversarial networks (GAN).
Mistral has released the Pixtral 12B
Mistral’s official account on X (formerly Twitter), known for its minimalist announcements, published the AI model in a post with a magnet link. The Pixtral 12B file size is 24GB in total and requires a PC equipped with an NPU or a computer with a powerful GPU to run the model.
The Pixtral 12B has 12 billion parameters and is based on the company’s existing AI model Nemo 12B. Mistral Highlight users also need a Gaussian Error Linear Unit (GeLU) as a vision adapter and a 2D-RoPE as a vision encoder.
In particular, users should be able to upload image files or URLs to Pixtral 12B and answer questions about the images, such as identifying objects, counting the number of objects, and sharing additional information. This Nemo-based model is also capable of handling all common text-based tasks.
A Reddit user posted a screenshot of the Pixtral 12B’s benchmark results, and it looks like the LLM outperforms the Claude-3 Haiku and Phi-3 Vision in multimodal capabilities on the ChartQA bench. It also outperforms both competing AI models in Massively Multitasking Language Understanding (MMLU) when it comes to multimodal knowledge and thinking.
Mistral’s AI models can be modified and used under the Apache 2.0 license, TechCrunch reported, citing a company spokesperson. This means that the results of the model can be used without restriction for personal or commercial purposes. Additionally, Sophia Young, head of developer relations at Mistral, announced in a post that the Pixtral 12B will soon be available on Le Chat and Le Platforme.
Currently, users can download AI models directly through the magnet link provided by the company. Alternatively, model weights are also hosted on the Hugging Face and GitHub lists.
Bhupendra Singh Chundawat is a seasoned technology journalist with over 22 years of experience in the media industry. He specializes in covering the global technology landscape, with a deep focus on manufacturing trends and the geopolitical impact on tech companies. Currently serving as the Editor at Udaipur Kiran, his insights are shaped by decades of hands-on reporting and editorial leadership in the fast-evolving world of technology.