This article was AI-generated as part of an experimental historical-content project. The date reflects the period being analyzed rather than the date the article was originally written.
Most of the ways a company gets described online still involve a server somewhere. A search engine crawls a page, ranks it and shows a snippet. A chatbot runs in a data center and sends an answer back. Reputation work has been organized around those central points, because that is where the descriptions were produced.
Meta’s Llama 3.2 release this week is a reminder that some of that production is moving.
Two different things in one release
Meta announced Llama 3.2 at its Connect event on Wednesday. It is worth separating the two halves of the release, because they are easy to blur together.
The first half is vision. Llama 3.2 includes 11-billion and 90-billion parameter models that accept images as well as text. Meta says they can handle document understanding, including charts and graphs, as well as image captioning and locating objects in an image from a description. As The Verge noted, these are Meta’s first open models that can process images. OpenAI and Google already had multimodal models. What is new is that developers can download, fine-tune and run these themselves.
The second half is size. Llama 3.2 also includes 1-billion and 3-billion parameter text-only models designed for phones and other edge devices, enabled on day one for Qualcomm and MediaTek hardware and optimized for Arm processors. Meta highlights summarization and rewriting as on-device uses, and says running locally means data such as messages and calendar information does not have to be sent to the cloud.
So the accurate summary is not “vision on the phone.” The vision models are the larger ones. The phone-sized models handle text. Together they make it much easier to build systems that look at images and summarize text without depending on a single large provider.
Why that matters for how brands get described
When a description of a company comes from Google or ChatGPT, there is at least one central system to test. You can run queries, record the answers and report errors through feedback channels. The descriptions are consistent enough to audit.
Open models that anyone can download change that. A developer can build a shopping app that looks at a product photo and describes it, a document tool that reads a chart in an annual report and summarizes it, or a phone assistant that condenses messages and articles about a company into one sentence. Each will make its own choices about prompts, fine-tuning and data. There is no single answer to check.
On-device processing adds another layer. If a summary is produced locally, from content a user already has, it never touches a server the company can observe. That is good for user privacy, which is the point. For brands it means some descriptions will form in places they cannot see at all.
Images also matter more than they used to. A model that can read a chart can misread one. A model that captions images will describe logos, packaging, storefronts and executives in photos. Visual assets that used to be interpreted mainly by people will increasingly be interpreted by software first.
What a reputation team can do
The practical response is not to chase every app. It is to strengthen the inputs most systems will share.
Put key facts in clear text, not only inside images or PDFs. If a chart in an investor presentation carries an important number, state the number in text nearby.
Treat official images as information. Captions, alt text and consistent naming help any system describe them correctly.
Keep the public record consistent across your own site, Wikipedia and major coverage, so that whatever a smaller model retrieves, or learned in training, points in the same direction.
None of this requires knowing which app will become popular. It requires accepting that the brand summary is no longer produced only in a few places you can name.