Meta unveils Llama 3.2 with vision and Orion AR glasses

On September 25, 2024, Meta introduced Llama 3.2 with vision models and lightweight on-device variants, and showed Orion as an augmented-reality glasses prototype. The Llama 3.2 documentation supports evaluation of a downloadable model; Orion belongs to another evidence category: research hardware not yet available to consumers.

The announcement brings together two strategies that have so far advanced along separate paths: models capable of interpreting the visual world and hardware that can display digital information directly over it.

Llama 3.2 gives Meta’s family eyes

Llama 3.2 comes in four models. The 11 billion- and 90 billion-parameter versions are multimodal: they can take text and images as input and respond with text. They can, for example, extract data from a chart, describe a photograph, answer questions about a scanned document or identify specific elements in an image. Source

A multimodal model is not simply a chatbot with a photo attached. It must connect what it sees with the user’s question and with language. This capability has become central to the competition among OpenAI, Google, Anthropic and Meta because it can automate tasks involving documents, screenshots, interfaces or product images.

Llama 3.2’s vision models are built on the Llama 3.1 architecture. Meta trained a visual layer so the model can translate information from an image into representations the language model can process. As with the rest of Llama, the company is releasing the weights under its community license: researchers and companies can download and adapt the model, subject to specific conditions for very large organizations.

That sets Llama apart from closed services such as ChatGPT and Claude. A company can install a model on its own infrastructure, exercise greater control over where its data remains and tailor it to a specific task. In return, it takes on the technical cost of deploying, evaluating and securing it.

1 billion- and 3 billion-parameter models for devices

The other new arrivals are Llama 3.2 1B and 3B, text-only models with 1 billion and 3 billion parameters, respectively. Their purpose is not to compete with the largest models on complex reasoning, but to make AI features viable on resource-constrained hardware. Source

Meta is positioning them for phones, personal computers and connected devices. Running some AI directly on the device — commonly known as edge AI — can reduce latency and keep each query from traveling to a remote server. That is a significant advantage for writing suggestions, information classification, voice commands or assistants that handle sensitive data.

The trade-off is unavoidable: a small model knows less and makes more mistakes than one with tens or hundreds of billions of parameters. In practice, the choice will not be between local AI and cloud AI, but how to divide the work. Simple, private actions are a better fit for the device; queries requiring more knowledge or computation will continue to rely on data centers.

Meta AI gets a voice, while Orion points to a phone-free future

Meta also announced new voices for Meta AI, its assistant built into WhatsApp, Instagram, Facebook and Messenger. The company is adding voice conversations and voices modeled on personalities including John Cena, Judi Dench, Awkwafina, Keegan-Michael Key and Kristen Bell, initially in the United States, Canada, Australia and New Zealand.

The event’s most ambitious demonstration was Orion. These are not smart glasses like the Ray-Ban Meta, which feature a camera, microphones and speakers, but augmented-reality glasses with transparent displays. They project digital elements into the user’s field of view, including messages, video calls, directions and contextual information.

The prototype uses Micro LED projectors, silicon carbide waveguides and eye- and hand-tracking. It can also be controlled with an electromyography-based wristband, a technology that detects the electrical signals from the wrist muscles even before the finger completes a gesture. Meta says Orion offers a field of view of about 70 degrees and weighs less than 100 grams. Source

Orion will not be available in stores for now. Meta is presenting it as a research platform: the hardware remains very expensive to manufacture, and the company needs to bring down the cost before commercializing it. Showing it nonetheless has strategic value. Meta wants to demonstrate that it has an alternative to phones and bulky mixed-reality headsets.

The challenge is turning demos into everyday utility

Llama 3.2 gives developers and companies an open way to experiment with vision and on-device AI. Orion, by contrast, highlights the distance that still separates a compelling demonstration from a consumer product.

The connection between the two announcements is clear: useful glasses will need to understand what the user is looking at, respond quickly and handle especially sensitive data. Privacy, battery life, visual-recognition accuracy and society’s willingness to wear cameras and displays on their faces will determine whether that vision makes it out of the lab. Meta has put the pieces on the table; turning them into an everyday device will require considerably more than a good language model.

Model, product and prototype require different tests

Llama weights allow inference to be repeated on owned hardware and memory, speed and quality to be measured. Meta AI adds an interface, policies and servers that cannot be inferred from those weights. According to the specific prototype announcement, Orion was offered to employees and selected audiences for learning; a controlled demonstration establishes neither battery life, price nor daily comfort.

For vision, separate description, text reading, charts and spatial grounding. Use blurred, cropped and contradictory images, and require the model to identify the region supporting its answer. A model may describe a scene and fail on the small number that changes a decision.

Local does not mean unexposed

Running a model on the device reduces transfers, but the application still has permissions, files and actions. Trace camera, microphone, storage and any escalation to cloud services. Glasses extend the issue to bystanders: a person in the field of view may not know what is being captured or processed.

The transferable skill is to label the maturity of every claim: available weights, deployed feature, limited beta, prototype or plan. Then apply the appropriate test to each state. This prevents a real model from lending credibility to hardware that remains experimental.

An evidence matrix organises the presentation

Create rows for capability, availability, price, autonomy, comfort and safety. In each, mark whether evidence comes from downloadable weights, public access, a vendor test or a future plan. Empty cells are not failures: they show what cannot yet be concluded.

Then design a test suited to each row. Reproduce a local model on described hardware; evaluate an assistant in available tasks and regions; test glasses through prolonged use outside a demonstration. Comparing everything through a promotional video removes the important differences.

Date belongs in the matrix. Labels such as beta and prototype can change, and an updated link should not rewrite what existed at announcement time. Preserving captures, version and conditions turns commercial memory into auditable evidence.

The matrix also prevents capabilities being added across objects. A vision model does not prove that glasses understand their environment, and a prototype displaying information does not guarantee that the assistant is available inside it.