Like every major technology company today, Meta has its own flagship model for generative AI, called lama. Llama is exclusive amongst major models in that it’s “open,” meaning developers can download it and use it nonetheless they need (with certain restrictions). This is in contrast to models like Anthropics' Claude, OpenAI's GPT-4o (which powers ChatGPT), and Google's Gemini, that are only accessible via APIs.
However, to provide developers selection, Meta has also partnered with providers equivalent to AWS, Google Cloud and Microsoft Azure to make cloud-hosted versions of Llama available. In addition, the corporate has released tools to make it easier to fine-tune and customize the model.
Here you'll find the whole lot it’s worthwhile to learn about Llama, from its features and editions to the way to use it. We'll keep this post updated as Meta releases upgrades and introduces latest development tools to support using the model.
What is Lama?
Llama is a model family – not only a model:
- Call 8B
- Call 70B
- Call 405B
The latest versions are Call 3.1 8B, Flame 3.1 70B And Call 3.1 405Bwhich was released in July 2024. They are trained using web pages in numerous languages, public code and files on the net, and artificial data (i.e. data generated by other AI models).
Llama 3.1 8B and Llama 3.1 70B are small, compact models designed to run on devices starting from laptops to servers. Llama 3.1 405B, however, is a large-scale model that requires (without some modifications) data center hardware. Llama 3.1 8B and Llama 3.1 70B are less powerful than Llama 3.1 405B, but faster. In fact, they’re “distilled” versions of 405B, optimized for low memory overhead and latency.
All Llama models have context windows of 128,000 tokens. (In data science, tokens are chunked raw data, just like the syllables “fan,” “tas,” and “tic” within the word “improbable.”) A model's context, or context window, refers to input data (e.g., text) that the model considers before generating output (e.g., additional text). An extended context can prevent models from “forgetting” the content of current documents and data and wandering off topic and drawing incorrect conclusions.
These 128,000 tokens correspond to roughly 100,000 words or 300 pages, which is roughly the length of Wuthering Heights, Gulliver's Travels, and Harry Potter and the Prisoner of Azkaban.
What can Lama do?
Like other generative AI models, Llama can perform quite a lot of support tasks, equivalent to coding and answering simple arithmetic questions and summarizing documents in eight languages (English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai). Most text-based workloads—equivalent to analyzing files like PDFs and spreadsheets—fall inside its purview; not one of the Llama models can process or generate images, although this is feasible. change within the near future.
All current Llama models will be configured to leverage third-party apps, tools, and APIs to get things done. They're able to use Brave Search to reply questions on current events, the Wolfram Alpha API for math-scientific queries, and a Python interpreter to validate code. In addition, in response to Meta, the Llama 3.1 models can use certain tools they haven't seen before (but whether or not they can use those tools is one other query).
Where can I exploit Llama?
If you only want to talk with Llama, it’s Supporting the Meta AI chatbot experience on Facebook Messenger, WhatsApp, Instagram, Oculus and Meta.ai.
Developers working with Llama can download, use, or optimize the model on hottest cloud platforms. Meta claims that over 25 partners host Llama, including Nvidia, Databricks, Groq, Dell, and Snowflake.
Some of those partners have built additional tools and services on top of Llama, including tools that allow the models to point to proprietary data and run them with lower latency.
Meta recommends using the smaller Llama 8B and Llama 70B models for general-purpose applications equivalent to chatbots and code generation. Llama 405B, the corporate says, is healthier suited to model distillation—the means of transferring knowledge from a big model to a smaller, more efficient model—and generating synthetic data for training (or fine-tuning) alternative models.
Important: The Llama License limits how developers can use the model: App developers with greater than 700 million monthly users must request a special license from Meta, which the corporate grants at its discretion.
In addition to Llama, Meta offers tools designed to make using the model “safer”:
- Call guard, a moderation framework
- Fast guard, a tool to guard against prompt injection attacks
- CyberSecEval, a cybersecurity risk assessment suite
Llama Guard attempts to detect potentially problematic content that’s either fed into or generated by a Llama model. This includes content related to criminal activity, child abuse, copyright infringement, hate, self-harm, and sexual abuse. Developers can adjust the categories of blocked content and apply the blocks to all languages that Llama supports by default.
Like Llama Guard, Prompt Guard can block text intended for Llama, but only text intended to “attack” the model and make it behave in undesirable ways. Meta claims that Llama Guard protects against explicitly malicious prompts (i.e. jailbreaks that try to bypass Llama's built-in security filters), in addition to prompts containing “Injected inputs.”
CyberSecEval is less a tool than a set of benchmarks for measuring model security. CyberSecEval can assess the chance a Llama model poses (no less than in response to Meta's criteria) to app developers and end users in areas equivalent to “automated social engineering” and “scaling offensive cyber operations.”
Lama’s limitations
Like all generative AI models, Llama comes with certain risks and limitations.
For example, it’s unclear whether Meta Llama trained on copyrighted content. If it did, users could possibly be responsible for copyright infringement in the event that they inadvertently use a copyrighted snippet that the model regurgitated.
Meta at one point used copyrighted e-books for AI training despite warnings from its own lawyers, as Reuters recently reported. The company controversially trains its AI using Instagram and Facebook posts, photos and captions and makes it difficult for users to deactivateIn addition, Meta, together with OpenAI, is the topic of an ongoing lawsuit by authors, including comedian Sarah Silverman, over the businesses' alleged unauthorized use of copyrighted data for model training.
It can also be advisable to be cautious when using Llama when programming. Because Llama – like its generative AI counterparts – could produce faulty or unsafe code.
As all the time, it’s best to have AI-generated code reviewed by a human expert before integrating it right into a service or software.
