What Meta Actually Opened With Llama 3 and Horizon OS

What Meta Actually Opened With Llama 3 and Horizon OS

Meta made two important announcements in April 2024, but the word open did too much work in the original conversation.

Llama 3 made model weights available for broad use under Meta's community license. Meta Horizon OS opened a mixed-reality platform to selected hardware partners. Those are meaningful forms of access. Neither announcement, by itself, meant that Meta had released everything as open-source technology.

That distinction is the durable story behind the episode.

Llama 3 was a real model release, not one model

Meta released pretrained and instruction-tuned Llama 3 models with 8 billion and 70 billion parameters on April 18, 2024. The company said larger models above 400 billion parameters were still training.

That timing matters. The episode discussed the larger model as forthcoming, not available. Meta later released Llama 3.1 405B in July 2024, but that later product should not be folded backward into what listeners could use on April 23.

Meta also said the Llama 3 training data contained more than 15 trillion tokens and four times as much code as the Llama 2 training set. The company attributed improved reasoning, code generation, instruction following, and lower false-refusal rates to a combination of data, architecture, pretraining, and post-training changes. It would be too simple to say that adding code alone caused the improvement.

A benchmark score is a result under a setup

The episode used MMLU to compare models. MMLU is a benchmark introduced by researchers to test language models across 57 academic and professional subjects, including mathematics, history, computer science, and law.

Calling it an SAT for language models is a useful first image, but it hides important details. MMLU measures answers to a particular collection of multiple-choice questions. A score can change with the model version, prompt format, number of examples supplied in the prompt, scoring method, and evaluation implementation.

Meta published strong benchmark results for the 8B and 70B models and linked to evaluation details. Those numbers were vendor-reported results, not a permanent league table. They supported the claim that Llama 3 was competitive in Meta's chosen evaluations. They did not prove that it was the best model for every task or that a smaller parameter count automatically meant a better product.

The practical question is still whether a model performs reliably on the work someone actually needs to do.

Meta AI turned the model into a product experience

The part of the episode that has aged best is Dalton's account of using Meta AI, not the exact speed ranking he gave it.

At launch, Meta put the Llama 3-powered assistant into Facebook, Instagram, WhatsApp, Messenger, and the web, with availability varying by country and feature. The company also introduced a beta version of its Imagine experience that changed generated images while a person typed.

Dalton used the assistant while setting up Venture Step's Facebook and Instagram presence. He found the conversational tone warmer than some alternatives and saw fewer refusals during that session. He was also impressed by the live image-generation interaction.

Those observations belong to a particular day, account, prompt, and product version. They are useful as firsthand product notes, not as a current comparison with Gemini, ChatGPT, or any other assistant.

Horizon OS opened a partner ecosystem

On April 22, Meta announced that third-party device makers would be able to build mixed-reality hardware on Meta Horizon OS. ASUS Republic of Gamers and Lenovo were named as device partners. Meta also described a limited-edition Quest collaboration with Xbox.

The operating system already powered Meta Quest headsets. The new strategy was to let more hardware makers use a common platform and to make it easier for developers to reach users through the renamed Meta Horizon Store.

That is an open-platform move. Meta did not say in the announcement that it was publishing the complete Horizon OS source code under an open-source license. The platform was built partly on the Android Open Source Project, but a product that includes open-source components is not automatically open source as a whole.

The difference matters because partners and developers depend on Meta's access terms, platform rules, store policies, identity layer, and technical roadmap. A wider ecosystem can create choice without becoming permissionless.

Open weights are not the same as open source

Meta repeatedly described Llama 3 as open source. The Open Source Initiative later concluded that the Llama 3.x community license did not satisfy its Open Source Definition because the license retained restrictions inconsistent with permission to use the system for any purpose.

The clearest description is that Llama 3 was an openly available or open-weight model under a Meta-authored license. People could download and build with the released weights, subject to that license and the rest of the operating requirements.

This is more than a vocabulary dispute. The label affects what a developer may do, who can use the model, whether outputs can support another model, how redistribution works, and what obligations attach to a product.

Access is the beginning of an adoption decision

Downloadable weights can reduce dependence on a single hosted interface. A common mixed-reality platform can reduce duplicated development work across devices. Both can give builders room to experiment.

Neither removes the need to examine the license, compute requirements, security controls, privacy path, safety tooling, evaluation evidence, platform governance, and long-term maintenance burden.

The most useful conclusion from Meta's announcements is not that open always beats closed. It is that different kinds of openness create different kinds of freedom and dependence. Builders should name the kind they are getting before they build a strategy around it.

Continue the conversation

The episode captures an immediate reaction to Llama 3, Meta AI's live image experience, and the Horizon OS partner announcement. Listen on Spotify or watch on YouTube.

Sources and editorial notes

The preserved [[E13 - Transcript - Google Drive recovered|raw source]] is a recovered production outline, not a verbatim transcript. This article uses that outline, the retained legacy episode record, Meta's April 2024 announcements, the original MMLU paper, Meta's later Llama 3.1 release, and the Open Source Initiative's license analysis. Current model availability, platform access, license terms, integrations, and partner plans require a fresh check before publication.