Meta releases Llama 3.1 with a 405-billion-parameter model

Meta launched Llama 3.1 today, a new family of language models available in 8-billion, 70-billion and 405-billion-parameter versions. The latter is, in terms of both size and capabilities, the largest publicly available collection of AI model weights to date.

The move puts Meta in a unique position against OpenAI, Anthropic and Google: it competes at the high end of the generative AI assistant market while allowing companies, researchers and developers to download the models and run or adapt them on their own infrastructure.

A 405B model competing in the top tier

Parameters are the internal values a model adjusts during training to recognize patterns and generate responses. They are not a perfect measure of quality, but they provide a sense of scale. Llama 3.1’s 405 billion parameters far exceed the 70 billion of the largest previous Llama 3 model.

Meta says Llama 3.1 405B competes with leading closed models on knowledge, mathematical reasoning, coding and instruction-following tasks. The company has published its results on standardized tests, although those comparisons should be read cautiously: benchmarks measure specific skills and do not fully reproduce how an assistant is used day to day.

The family includes a context window of up to 128,000 tokens. In practical terms, it can handle long documents, extended conversations or large blocks of code without losing the thread as quickly. It also improves tool use: a developer can configure the model to query a database, call an API or run a function instead of simply generating text.

The models are trained to work in eight languages: English, German, French, Italian, Portuguese, Hindi, Spanish and Thai. For Spanish-speaking users, the fact that a family of this scale includes Spanish from launch matters more than an interface translation: it affects how well the system understands instructions, writes and handles local texts.

Weights available, but not a license without conditions

Meta distributes Llama 3.1 under its community license. This allows the model to be used, modified and redistributed for many commercial and research projects. However, calling Llama simply “open source” requires an important qualification.

The company publishes the weights—the result of the training—but has not disclosed the full dataset or every detail needed to reproduce the creation process from scratch. Its license also imposes conditions: companies with more than 700 million monthly active users must request specific authorization from Meta to use Llama.

Llama 3.1 is therefore far more accessible than a model available only through a paid API, but it is not equivalent to free software in the strictest sense. That distinction matters to organizations seeking technological independence or needing to audit a system’s origins precisely.

Zuckerberg’s case for open AI

The launch is accompanied by a statement from Mark Zuckerberg defending open AI. Meta’s CEO argues that distributing models reduces the concentration of power, makes it easier for more companies to build their own products and improves safety by allowing outside researchers to examine the technology.

It is also a business strategy. Meta does not rely on selling access to a single chatbot as its primary source of revenue. Its business is built around advertising and social platforms, so it can absorb the cost of publishing models to attract developers, encourage tools built around Llama and reduce its dependence on rival providers.

The price of that openness is that advanced capabilities reach more hands. Meta has trained and fine-tuned its models to reject harmful requests, but downloadable models give users more control than a centralized service. The debate over whether that distribution improves collective safety or increases risks will continue to grow as open models approach the best closed systems.

What changes for companies and developers

The 405B model will not be easy to deploy internally. Running it requires substantial and expensive GPU infrastructure, especially if the goal is to serve many users at once. For most companies, the 8B and 70B models will be more realistic options: they can be specialized with proprietary data, run at lower cost and, in some cases, remain within the organization’s own infrastructure.

Meta has announced that Llama 3.1 will be available through cloud providers and platforms including AWS, Microsoft Azure, Google Cloud, NVIDIA, Databricks and IBM. That distribution lowers the barrier to testing it without buying hardware, although it does not eliminate computing costs.

The main development is not that everyone will download a 405-billion-parameter model onto their computer. It is that small labs, technology companies and infrastructure providers can now study, fine-tune and turn a model of this scale into products. Meta has abruptly raised the bar for what can be built outside the closed services of the major AI companies.