Meta releases Llama 3, brings Meta AI to WhatsApp and Instagram

On April 18, 2024, Meta released the first Llama 3 models—8-billion and 70-billion-parameter variants—and expanded Meta AI inside its applications. Meta’s announcement documents training, evaluations, licence and initial rollout; “openly available” describes conditional access rather than unrestricted free software.

The first release includes models with 8 billion and 70 billion parameters. Parameters are the values a model adjusts during training to learn patterns in language, code and other data. Meta also confirmed that it is training a much larger version with more than 400 billion parameters, although it has not set a release date. Primary source.

Two open models with a proprietary license

Llama 3 comes in base versions for those who want to adapt them, as well as versions fine-tuned for conversation and following instructions. Meta says the 70 billion-parameter model represents a significant improvement over Llama 2 in reasoning, coding and instruction following, and competes with similarly sized commercial models on several benchmarks. Primary source.

The company trained this generation on more than 15 trillion tokens, the units of text models process. According to Meta, that figure is more than seven times the amount of data used for Llama 2. It has also increased the share of coding data and content in languages other than English. Primary source.

Llama’s openness does not mean it is open-source software in the strictest sense. Its weights—the files needed to run the model—can be downloaded and used for research and commercial purposes under Meta’s community license. However, companies with more than 700 million monthly active users must request specific authorization. Primary source.

That distinction matters because Llama has become one of the most widely used alternatives to closed models from OpenAI, Google and Anthropic. Being able to download it lets organizations run it on their own servers, adapt it to a specific task and avoid routing every query through an external API. In return, whoever deploys it takes on responsibility for the infrastructure, security and evaluation of its potential failures.

Meta AI enters the company’s apps

The most visible novelty for the public is not the model but the assistant that uses it. Meta AI is beginning to roll out in English in the United States within WhatsApp, Instagram, Facebook and Messenger, as well as on the meta.ai website. The assistant can answer questions, help plan tasks, recommend options and generate images from written prompts.

In some Facebook and Instagram search features, Meta AI will appear directly in the box where users already type their queries. The company wants AI to be not a separate app users have to open deliberately, but a layer built into the places where they already chat, search for content and share links.

Meta AI can also draw on web search results provided by Bing. That is a significant decision: on its own, a language model responds based on the information it learned during training and may be outdated or wrong. With web access, its answers can incorporate recent information, although they still need to be checked when important matters are involved.

The challenge is not just technical

Meta has a distribution advantage that is difficult to match. WhatsApp, Instagram, Facebook and Messenger bring together billions of people, and the assistant is being introduced into products that are already part of their routines. OpenAI popularized the chatbot as a standalone destination with ChatGPT; Meta is trying to make conversations with AI happen inside existing social and messaging apps.

But that advantage raises practical questions. In a group conversation, for example, invoking Meta AI means sending a request to the company’s systems—something worth keeping in mind before sharing sensitive data. Meta will need to clearly explain what information it uses to improve its services, how it separates private conversations from queries to the assistant and how it prevents incorrect answers in products used at massive scale.

For now, the rollout is limited to English and the United States. The decisive test will be whether Meta AI proves useful beyond image-generation demos and quick answers, and whether Llama 3 can establish itself as a technical foundation for companies and developers that do not want to depend on a single provider of closed models.

Weights, licence and assistant are not the same

Downloadable Llama 3 supports local execution and adaptation; Meta AI adds search, interface, policies and private systems. An assistant output does not necessarily describe the weights, and a local test does not reproduce the product. The record should name variant, tuning and tool layer.

The community licence allows broad uses with conditions and a threshold for very large services. Preserve the text revision, check attribution duties and separate trademark from licence before distribution. “Open” without those details may hide an important restriction.

How to compare variants

Size affects memory, latency and quality but does not select a task by itself. Test classification, extraction, writing and code with in-house cases and assign the smallest model that clears the threshold. Always using the largest wastes resources; using the smallest without review transfers cost to errors.

Declared multilingual data do not guarantee parity. Write original tests in each language with dialects, formats and local knowledge, and inspect language mixing. An English improvement may coexist with a regression where the product is deployed.

Distribution inside messaging amplifies both utility and error. Users should distinguish a human conversation from Meta AI, know when an answer uses the web and be able to cite or correct its source. Integration in a familiar application reduces friction; it may also reduce the critical pause before belief or sharing.

The transferable skill is to audit a family by artefact, licence, variant, language and channel. That inventory uses Llama 3 without confusing weight access with complete transparency or application presence with reliability.

Tuning defines visible behaviour

A base variant predicts continuations; an Instruct variant has received examples and preferences for following requests. Preserve both labels in comparisons. Fine-tuning on in-house data may improve vocabulary while degrading safety or general knowledge, so repeat earlier tests alongside the new task.

Meta’s published evaluation set also involved human and model participation. Opening categories and criteria reveals what was rewarded. Aggregate preference does not replace factual accuracy, especially when the assistant retrieves web information.

Check the exact licence and model revision before integration. Families change and a short name may cover different windows, tuning or filters. A hash and provenance record reconstruct the system after an incident.

Updates need regression tests. A new variant may improve coding and worsen a language or data format. Shipping solely because the average rose treats particular users as statistical noise.

Model documentation should accompany deployment inside applications. When Meta AI uses another layer or changes version, users need a signal. Traceability prevents results from published weights being used to justify a service nobody can inspect in the same way.

The chosen model must be removable when it fails the threshold.