Meta published Llama 4.2 on Friday across four sizes, and the interesting one is the smallest. The 3 billion parameter variant is distilled specifically for the neural coprocessor in the Ray-Ban Display glasses, running quantized to 3 bits with a custom kernel that Meta released alongside the weights. It handles translation, scene description, and short question answering entirely on the glasses, with no phone and no network. Latency to first token measured 310 milliseconds on Meta's hardware. The larger 90 billion and 405 billion parameter models improve modestly on the 4.1 release, roughly two points on most reasoning benchmarks, which is the expected shape of a point release.
What running on glasses requires
The Ray-Ban Display's coprocessor has 2 gigabytes of memory shared with everything else the device does, a power budget measured in hundreds of milliwatts, and a thermal ceiling set by the fact that it sits against a human face. A 3 billion parameter model at 3-bit quantization occupies about 1.3 gigabytes, which leaves room for the KV cache and nothing else. Meta's kernel work focuses on keeping weights resident and streaming activations, since the memory bandwidth is the binding constraint rather than compute.
Quality at 3 bits is the obvious concern, and Meta addressed it with quantization-aware distillation rather than post-training compression. The 3B model was trained with the quantization in the loop, which costs more to produce and preserves substantially more capability. On the translation tasks Meta targets, the quantized model scores within 1.4 BLEU of the full precision version. On general reasoning it degrades much more, which Meta acknowledges by scoping the on-device use cases narrowly.
The features this enables
Live translation without connectivity is the flagship use case and the one Meta demonstrated. Speaking Spanish to someone wearing the glasses produces English text in the display with about 900 milliseconds of lag, and it works in an airplane, a subway, or a rural area with no signal. Meta supports 22 language pairs on device, against 47 through the cloud path. The difference in quality is noticeable on idiomatic speech and negligible on straightforward exchanges.
Scene description for low-vision users is the application with the most obvious social value and it works better than I expected. Asking what is in front of me produces a two-sentence description in about 1.2 seconds. The Royal National Institute of Blind People has been testing an early build and its technology lead said the offline capability matters enormously, because the situations where a blind user most needs assistance are often the ones with the worst connectivity.
The story is rarely the launch. It is what breaks, what ships, and who owns the mess at 2 a.m.
The licensing situation
Llama 4.2 ships under the same community license as previous versions, which permits commercial use below 700 million monthly active users and requires attribution. That license is not open source by the Open Source Initiative's definition and Meta has stopped arguing the point. The 3B model carries an additional restriction: the quantization kernels are released under a separate license that permits use on any hardware and prohibits redistribution in a modified form for competing wearables, which several developers described as an unusually specific constraint.
The practical effect is that Meta will happily let anyone run the 3B model on a phone or a laptop and has made it awkward to build a competing pair of glasses around it. That is a coherent commercial position and it complicates the story Meta tells about open models. Researchers at the Allen Institute pointed out that the kernel license makes independent evaluation on non-Meta hardware harder, since the reference implementation is what everyone would benchmark against.
How this compares to what Apple and Google are doing
Apple runs a roughly 3 billion parameter model on iPhone for its intelligence features and has never published weights or detailed architecture. Its on-device model handles a similar range of tasks and benefits from tighter hardware integration and a much larger memory budget on a phone. Google's Gemini Nano runs on Pixel and on Samsung flagships and has a public API, with weights unavailable.
Meta is alone in publishing weights for a model at this scale that is genuinely tuned for a constrained device, which matters for research and for the long tail of hardware nobody at a large company is thinking about. A team building a medical device, an industrial wearable, or an assistive tool now has a starting point that does not require negotiating with a platform vendor. That is the most consequential thing about this release and it is not what Meta emphasized in its announcement.
What comes next
Meta said a vision-language variant of the 3B model arrives in the autumn, which is what the scene description feature currently approximates through a separate vision encoder plus the language model. Merging them would cut latency and memory. The company also confirmed work on a model that runs on the audio processor in ordinary Ray-Ban glasses without a display, targeting the far larger installed base.
The competitive question is whether Meta's hardware advantage compounds. Roughly 4 million Ray-Ban Meta units have shipped since 2023 according to IDC estimates, which is small in phone terms and enormous for a smart glasses category that has failed repeatedly. Each generation adds capability that requires the model work Meta is publishing, and publishing it recruits developers who improve it. That flywheel is the strategy, and it is working better than Meta's metaverse spending ever did.
Skarvonix will keep following this beat with reporting grounded in how systems behave outside the launch keynote.
- LLMs
- Open Source
