What Is Llama? Meta's AI Models Explained
Llama is Meta's family of large language and multimodal AI models whose weights can be downloaded and deployed by developers under Meta's community licenses. Organizations can use Llama through a hosted API, a cloud provider or infrastructure they operate themselves.
That makes Llama different from a consumer assistant such as ChatGPT or Meta AI. Llama is a model layer. It does not automatically include a polished chat interface, web search, account memory, connected apps, business administration or a safety system for a particular use case. Those capabilities are assembled around the model.
The current public flagship generation is Llama 4, released in April 2025 and documented in Meta's official Llama 4 announcement. Meta made two instruction-tuned models available:
- Llama 4 Scout, a 17-billion-active-parameter mixture-of-experts model with 109 billion total parameters and a stated context window of up to 10 million tokens;
- Llama 4 Maverick, a 17-billion-active-parameter mixture-of-experts model with 400 billion total parameters and a stated context window of up to one million tokens.
Both accept text and images and generate text or code. Both are large, datacenter-class models despite having only 17 billion parameters active for a token. Smaller Llama 3.x models remain more realistic for laptops, edge devices and low-cost local experimentation.
Llama at a glance
| Question | Short answer |
| What is Llama? | Meta's open-weight family of language and multimodal AI models |
| Is it a chatbot? | Not by itself; applications and chat interfaces are built around it |
| What is the latest public generation? | Llama 4 |
| Which Llama 4 models are downloadable? | Scout and Maverick |
| What can Llama 4 process? | Text and images as input, with text and code output |
| Is Llama open source? | Meta calls it open, but the weights use a custom community license rather than an OSI-approved open-source license |
| Can Llama be used commercially? | Generally yes, subject to the applicable license, acceptable-use policy and additional terms |
| Can Llama run locally? | Yes, but model size matters; smaller Llama 3.x models are the practical laptop route |
| Is Llama the model behind Meta AI? | It powered the 2025 Meta AI launch, but current first-party Meta AI uses Muse models |
| Is there a Llama API? | There is no single universal production endpoint: Meta launched a limited preview in 2025, while clouds, inference providers and self-hosting provide current routes |
| Does Meta train on Llama API prompts? | Meta said its 2025 limited preview did not use prompts or responses for training; current terms depend on the actual provider and service |
| What is Llama best for? | Custom AI applications, private deployment, model research, fine-tuning and supplier portability |
Why Llama matters
Most frontier AI products are closed services. A developer sends data to a provider's endpoint and receives a result, but cannot download the underlying model weights.
Llama offers a different proposition. The weights can be downloaded after accepting the applicable license. That creates several possibilities:
- deploy inside a private cloud or on-premises environment;
- select the hardware and inference runtime;
- fine-tune or adapt the model;
- use a specialist host and move later;
- build an offline or edge application with a smaller model;
- inspect the model card, tokenizer and deployment components;
- and avoid tying an application permanently to one API endpoint.
This is why Llama is important even when another model wins a benchmark. It changes the ownership and deployment model for AI infrastructure.
Open weights do not remove costs. The organization still needs compute, model operations, security, monitoring, evaluation, legal review and a product layer. A closed API can be cheaper and safer for a small workload because the provider absorbs those tasks.
Llama is not Meta AI, Muse or one single model
Meta operates several AI layers.
Meta AI is a product
Meta AI is the assistant in the Meta AI app, meta.ai and supported Meta apps and glasses. It includes interface, search, personalization, media tools and connected experiences.
Muse is Meta's current proprietary first-party model family
Muse Spark now powers the main Meta AI assistant. Muse Spark 1.1 is also offered through the Meta Model API, and Muse Spark 1.2 is available through that API and Muse Code at the August 2026 cut-off.
Llama is the open-weight developer family
Llama's purpose is broader than Meta's own consumer assistant. Developers can build their own chatbots, document systems, coding tools, agents, classifiers and multimodal applications around the weights.
A product may use multiple models
An assistant can route text, images, search, moderation and speech through different systems. “Which model powers this app?” may not have one permanent answer. Meta AI's move from Llama 4 to Muse demonstrates why product and model pages need separate update cycles.
The Llama model timeline
LLaMA and Llama 2: research becomes a platform
Meta first released LLaMA as a research-oriented model family. Llama 2 expanded commercial access and made the model ecosystem relevant to startups and enterprises, not only academic researchers.
The core proposition was already visible: provide capable weights that can be adapted and hosted outside Meta.
Llama 3: stronger general-purpose models
Llama 3 improved language quality and became widely available in 8B and 70B sizes. These models helped standardize local AI workflows because the smaller checkpoint could run on high-end consumer hardware after quantization, while the larger model suited servers and cloud GPUs.
Llama 3.1: scale and long context
Llama 3.1 introduced 8B, 70B and 405B models with a 128,000-token context window. The 405B model demonstrated that Meta intended Llama to compete at the frontier while preserving a downloadable-weights strategy.
Llama 3.2: small and vision models
Llama 3.2 split into two useful branches:
- 1B and 3B text models for smaller devices and local use;
- 11B and 90B vision models that accept images as well as text.
For many people asking how to run Llama locally, the 1B and 3B models are still more relevant than Llama 4. They fit ordinary hardware far more easily.
Llama 3.3: a capable 70B text model
Llama 3.3 focused on a 70B instruction-tuned text model. It remains useful where a mature text-only ecosystem, broad provider support and lower serving complexity matter more than Llama 4's native multimodality.
Llama 4: multimodal mixture of experts
Llama 4 introduced two major changes:
- native multimodality, combining text and image processing in one model backbone;
- mixture-of-experts architecture, activating only part of the total parameter set for each token.
Meta's Llama 4 model card is the primary specification source. It lists Scout and Maverick as instruction-tuned models intended for multilingual chat, visual reasoning and related commercial or research uses.
Llama 4 Scout versus Maverick
| Characteristic | Llama 4 Scout | Llama 4 Maverick |
| Active parameters | 17B | 17B |
| Total parameters | 109B | 400B |
| Routed experts | 16 | 128 |
| Stated context window | Up to 10M tokens | Up to 1M tokens |
| Input | Text and images | Text and images |
| Output | Text and code | Text and code |
| Typical role | Long-context, efficient multimodal workloads | Higher-quality general assistant, reasoning and visual workloads |
| Hardware class | Datacenter; Meta says an int4 configuration can fit on one H100 | Datacenter; Meta says an FP8 configuration fits on one H100 DGX host |
These numbers need interpretation.
Active parameters are not total memory
A mixture-of-experts model routes a token through only a subset of its parameters. This can reduce computation per token. It does not mean the inactive experts disappear from storage or memory. The complete model still needs to be available to the serving system.
A “17B active” Llama 4 model is therefore not equivalent to a conventional dense 17B model for download size or hardware planning.
Context length is a ceiling, not a quality guarantee
A stated 10-million-token window is technically significant, but an application should not assume perfect recall or reasoning across every token. Long-context performance depends on:
- prompt structure;
- document quality;
- retrieval method;
- model host limits;
- available memory;
- and the specific information the model must find.
Third-party providers may expose shorter limits than Meta's theoretical model maximum.
Scout is not a normal laptop model
The word “Scout” can sound small. It is smaller than Maverick, but it still has 109 billion total parameters. Meta's own hardware statement refers to an NVIDIA H100 with int4 quantization, not a typical consumer laptop.
For local use, start with our Llama local-deployment guide rather than downloading the largest model first.
How does mixture of experts work?
A dense model applies most or all model parameters to each token. A mixture-of-experts model contains multiple specialized feed-forward networks and uses a router to select the relevant expert path.
A simplified process is:
- the model receives a token and its context;
- a routing layer scores the available experts;
- selected experts process the token;
- their output is combined with shared layers;
- the model predicts the next token.
The advantage is more total capacity without using every parameter for every token. The trade-offs include routing complexity, memory movement, distributed serving and the need for optimized kernels and networking.
For a deeper infrastructure view, read AI World Today's guide to AI GPUs and AI infrastructure explainer.
What does native multimodality mean?
Llama 4 uses an early-fusion approach that integrates text and vision tokens into one model backbone. In practice, a Llama 4 application can accept an image and text instruction together, then return text or code.
Possible uses include:
- describing and classifying images;
- answering questions about diagrams;
- extracting information from screenshots;
- comparing product photos;
- generating code from a user-interface image;
- and combining a document image with textual context.
The model card says the released Llama 4 models were tested for image understanding with up to five input images. Developers using additional images or untested workflows need their own evaluation and safeguards.
Native multimodality does not provide native image generation. Llama 4 understands images and produces text; a separate image model is needed to generate pixels.
Which Llama model should you choose?
The newest model is not automatically the right model.
Choose Llama 4 Maverick when
- high-quality multimodal understanding is central;
- you have datacenter-class inference capacity;
- a hosted provider supports the required context and features;
- general assistant quality matters more than minimum cost;
- or visual reasoning is a core workload.
Choose Llama 4 Scout when
- exceptionally long context is valuable;
- you need a more efficient Llama 4 option;
- multimodal input is required;
- and the deployment can still support a 109B-total-parameter model.
Choose Llama 3.3 70B when
- text-only work dominates;
- provider and tooling maturity matter;
- Llama 4 is unnecessarily large;
- or the application has already been evaluated on the 70B model.
Choose Llama 3.2 1B or 3B when
- local or edge deployment is the goal;
- latency and memory matter more than frontier quality;
- the task is narrow;
- or you are learning the local Llama workflow.
Choose a third-party fine-tune when
- the base model is not specialized enough;
- the provider publishes clear provenance and license information;
- and you can test whether the fine-tune actually improves the target task.
Do not choose a model only from parameter count. Evaluate representative tasks, output quality, latency, memory, cost, safety and operational support.
How can you access Llama?
1. Download the weights from Meta or an approved host
Meta's model-download documentation explains how to obtain models directly or through Hugging Face and Kaggle after accepting the license.
This route provides maximum control but also maximum responsibility.
2. Use a hosted Llama endpoint
Meta announced a limited free preview of Llama API at LlamaCon in April 2025. The preview included API keys, playgrounds, Python and TypeScript SDKs, OpenAI SDK compatibility, Scout and Maverick access, evaluation tools and selected fine-tuning capabilities. Meta also said that prompts and responses in that preview were not used to train its models.
That launch announcement should not be treated as proof that every developer now has one permanent, generally available Meta production endpoint. In 2026, Meta's public first-party developer emphasis is the separate Meta Model API for proprietary Muse models. Llama can also be accessed through clouds, specialist inference providers and infrastructure the customer operates.
Our Llama API guide explains how to choose among those routes and how to write portable client code without inventing a universal base URL.
3. Use a cloud or inference provider
Llama is available through multiple clouds and specialist inference services. Benefits can include:
- serverless billing;
- managed scaling;
- regional deployment;
- enterprise contracts;
- optimized inference;
- and access to several model families through one API.
The provider's context limit, safety layer, retention, logging and pricing can differ from Meta's model card.
4. Run a smaller model locally
Tools such as Ollama, LM Studio and llama.cpp make smaller or quantized Llama models accessible on desktops and workstations. Local use can improve privacy and offline availability, but the user is responsible for updates, access control and model behavior.
Is Llama open source?
“Open-weight” is the most precise general description.
Meta publishes weights and supporting material under a custom Llama Community License. The Llama 4 Community License grants broad rights to use, reproduce, modify and distribute, including commercial use, but it also imposes conditions that standard permissive open-source licenses do not. The Open Source Initiative has argued that Meta's community-license restrictions do not satisfy the Open Source Definition.
Examples include:
- including the license when redistributing Llama materials;
- displaying “Built with Llama” for covered products or services;
- preserving an attribution notice;
- naming a distributed model beginning with “Llama” when Llama materials or outputs were used to create or improve it;
- obtaining a separate license when the organization exceeded 700 million monthly active users at the release threshold;
- following an acceptable-use policy;
Our dedicated guide explains what the Llama license actually allows.
Is Llama safe?
A model is only one component of a safe system.
Meta's model card recommends system-level safeguards and references tools such as Llama Guard, Prompt Guard and Code Shield. Developers still need to evaluate the deployment for:
- hallucinations;
- prohibited or harmful output;
- prompt injection;
- data leakage;
- insecure code;
- model-supply-chain risk;
- unauthorized access;
- and misuse of tools or external systems.
A locally hosted model is not automatically safer. It can keep prompts inside a controlled environment, but the organization becomes responsible for every operational control the cloud provider would otherwise supply.
For consumer and business risk across Meta AI and Llama deployments, read Is Meta AI safe?.
Llama versus DeepSeek and Gemini
Llama's strategic position becomes clearer when compared with two different alternatives.
Llama versus DeepSeek
Both ecosystems publish downloadable model weights, but their licenses, hosted services, model sizes, geography and API economics differ. DeepSeek currently emphasizes very low-cost hosted V4 models and MIT-licensed weights, while Llama has a broader Western cloud ecosystem and a more restrictive custom license.
See Llama versus DeepSeek for the full deployment comparison.
Llama versus Gemini
Gemini is primarily a proprietary managed-model ecosystem delivered through Google AI Studio, Gemini API, Vertex AI and Google products. Llama offers more deployment control; Gemini offers a more integrated managed platform and current multimodal model catalogue.
See Llama versus Gemini for the model, price and infrastructure decision.
How to evaluate a Llama deployment
Define the layer
Are you evaluating a base model, an instruction tune, a third-party fine-tune, a hosted endpoint or a complete application? Results from one layer do not transfer automatically to another.
Record the exact configuration
Capture:
- model name and revision;
- quantization or precision;
- runtime;
- context limit;
- prompt template;
- system prompt;
- decoding settings;
- hardware;
- and safety components.
Use task-specific tests
A general leaderboard cannot tell you whether the model extracts the correct fields from your contracts or produces code that passes your tests.
Measure total cost
Include:
- GPU time or API tokens;
- idle capacity;
- networking;
- engineering;
- monitoring;
- retries;
- human review;
- and failed outputs.
Test failure modes
Include malicious documents, conflicting instructions, long inputs, missing data, unsupported languages and attempts to reveal system information.
Plan an exit
One advantage of open weights is portability. Preserve prompts, evaluations, adapters and application contracts so the model can be replaced without rebuilding the entire product.
Frequently asked questions
What does Llama stand for?
The original name was styled LLaMA, for Large Language Model Meta AI. Meta now generally uses the brand name Llama rather than spelling out the acronym.
Is Llama a chatbot?
No. Llama is a model family. A chatbot can be built around Llama by adding an interface, system prompt, retrieval, tools, memory and safety controls.
Is Llama free?
The weights can be used without a conventional per-token royalty under the applicable license, but hosting is not free. APIs charge for usage, and self-hosting requires hardware and operations.
What is the latest Llama model?
The latest publicly released flagship generation at the August 7, 2026 cut-off is Llama 4, with Scout and Maverick weights available. Meta may release newer models later, so check its official repository.
Is Llama 4 better than Llama 3?
Llama 4 adds native multimodality, mixture-of-experts architecture and much longer stated context. Llama 3.x can still be better for smaller hardware, mature text-only deployments or lower operational complexity.
Can Llama 4 run on a laptop?
Not realistically in its full official form on an ordinary laptop. Scout and Maverick are datacenter-class. Smaller Llama 3.2 models and quantized Llama 3.x checkpoints are the practical local options.
Can I use Llama commercially?
Generally yes, but commercial use is subject to the exact version's community license, acceptable-use policy and additional restrictions. Large platforms and EU-based users of Llama 4 multimodal weights require particular attention.
Is Llama open source?
It is open-weight, but the custom community license is not an OSI-approved open-source license. The distinction matters for redistribution, attribution and certain users.
Does Meta AI still use Llama?
Meta AI used Llama 4 at its 2025 app launch. Meta says its current first-party assistant is powered by Muse Spark, with newer features driven by Muse Spark 1.1. Llama remains Meta's separate developer model family.
Is Llama safer when self-hosted?
Self-hosting can keep prompts under the operator's control, but it transfers security, moderation, monitoring, patching and incident response to that operator. Safety depends on the complete system.
The bottom line
Llama is Meta's open-weight AI infrastructure strategy. It gives developers more control than a closed API because the model can be downloaded, adapted and deployed across different environments.
Llama 4 Scout and Maverick bring native multimodality, mixture-of-experts architecture and very long context windows, but they are not small consumer models. The broader Llama family remains valuable precisely because it spans datacenter deployments and compact Llama 3.2 models that can run locally.
The right question is not simply “Which Llama is newest?” It is: which model, license, hosting route and safety architecture produce the best accepted result for this workload?
For the consumer assistant, app surfaces and current Muse model stack, return to the complete Meta AI guide.
byDavid Porter
Popular news
Anthropic Researchers Warn Advanced AI Models Could Trigger Human Extinction
Senator Josh Hawley Launches Massive Investigation Into OpenAI After Critical Hugging Face Breach
Andy Burnham Faces Urgent Demands to Block Artificial Superintelligence in Manchester
UK Computer Science Graduates Face a Massive Job Market Shift Driven by AI
Anthropic Researchers Issue Urgent Warning Regarding Elon Musk and Global AI Safety Risks
OpenAI Board Member Paul Christiano Warns AI Control Could Be Lost Forever
Latest comments
Loading