Llama

Also known as: Meta Llama

On this page 7

Llama (originally styled LLaMA, also called Meta Llama) is a family of large language models released by Meta with downloadable weights, so organizations can run, inspect and fine-tune them on their own hardware instead of reaching them only through a hosted API. Among named AI models and assistants in artificial intelligence, Llama is the reference point for the open-weight ecosystem: much of the tooling for local inference, quantization and fine-tuning was first built around it.

The family spans several generations since 2023, from small models that run on a laptop to frontier-scale systems that need a multi-GPU cluster. Llama models also power Meta AI, the assistant built into Meta’s consumer apps, but the defining feature of the family is that the weights themselves are distributed under a custom license.

How the Llama model family evolved

Llama evolved from a research release in 2023 into a broad product line with commercial rights, longer context and multimodal input. The first LLaMA models were offered to researchers under a non-commercial license, and their weights quickly circulated widely, which triggered a wave of community fine-tunes. Each later generation widened both the license terms and the range of sizes.

The following table summarizes the main generations and what each one changed.

GenerationReleasedSizesMain change
LLaMAFebruary 20237B, 13B, 33B, 65BResearch-only license, strong results for its size
Llama 2July 20237B, 13B, 70BCommercial use allowed, chat-tuned versions
Llama 3 and 3.1April and July 20248B, 70B, 405BAbout 15 trillion training tokens, 128K context
Llama 3.2 and 3.3September and December 20241B, 3B, 11B, 90B, 70BSmall edge models and image-input variants
Llama 4April 2025Scout and Maverick, 17B activeMixture-of-experts design, native multimodality
Main Llama generations and their key changes

Parameter counts such as 8B or 70B refer to billions of weights. For the mixture-of-experts Llama 4 models, the active parameter count per token is much smaller than the total, which reduces compute per token but not the memory needed to hold all experts.

Architecture behind Llama models

Llama models are decoder-only Transformers trained to predict the next token, refined with a set of design choices that later became common across open models. These choices improve training stability, inference speed and the handling of long inputs without changing the basic Transformer recipe.

FeatureWhat it doesPractical effect
RMSNorm pre-normalizationNormalizes inputs before each sublayerMore stable training at scale
SwiGLU activationGated feed-forward nonlinearityBetter quality per parameter
Rotary position embeddingsEncodes token position by rotating vectorsSupports context extension techniques
Grouped-query attentionShares key and value heads across query headsSmaller memory cache, faster inference
Mixture of experts (Llama 4)Routes each token to a subset of expert layersLower compute per token at large scale
Key architectural features of Llama

Each generation ships as a base model, trained only on raw text, and an instruction-tuned model, further trained with supervised fine-tuning and preference optimization to follow requests and refuse harmful ones. Developers who want to build their own assistant behavior usually start from the base model; those who want a ready chat model start from the instruct version.

Llama license and the open-weight debate

Llama is open-weight, not open source in the strict sense defined by the Open Source Initiative. The Llama Community License allows free use, modification and redistribution, but it includes an acceptable use policy, attribution requirements and a clause requiring a separate license from Meta for companies whose products exceeded 700 million monthly active users at release time.

The training data and full training code are not published, which limits reproducibility and independent auditing. Some releases also carried regional restrictions on multimodal versions. Legal teams should therefore review the specific license version attached to each model, because terms differ between generations.

Llama compared with other open-weight families

Llama competes with several other families of downloadable models that occupy similar sizes and use cases. Differences lie mainly in license terms, language coverage and size range, while benchmark rankings shift with almost every new release.

FamilyDeveloperTypical licenseNotable strength
LlamaMetaCustom community licenseLargest tooling ecosystem and fine-tune base
MistralMistral AIApache 2.0 for many modelsEfficient models, permissive terms
QwenAlibabaApache 2.0 for most sizesStrong multilingual and coding results
GemmaGoogleCustom Gemma termsCompact models for single-device use
gpt-ossOpenAIApache 2.0Reasoning-oriented open-weight models
Llama and comparable open-weight model families

Running and adapting Llama in practice

Running Llama is mainly a question of memory, quantization and serving software. At 16-bit precision each parameter needs two bytes, so an 8B model requires about 16 GB of memory for weights alone, while a 70B model needs around 140 GB. Quantization to 4 bits reduces these figures roughly fourfold at a modest quality cost, which is why tools such as llama.cpp and Ollama made local use practical.

SituationRecommended approachReason
Prototype on a laptop8B or smaller, 4-bit quantizedFits in consumer memory
Private enterprise chatbot70B instruct on dedicated GPUsGood quality, data stays in-house
Domain-specific taskLoRA fine-tune of a mid-size modelLow cost, targeted improvement
High-volume API serviceOptimized server such as vLLMBatching gives high throughput
Choosing a Llama deployment approach

Parameter-efficient fine-tuning, especially LoRA and its quantized variant QLoRA, lets teams adapt Llama to internal terminology on a single GPU. However, fine-tuning can weaken safety behavior learned during instruction tuning, so adapted models need fresh evaluation before release.

  • Do not assume benchmark scores transfer to a specific domain without testing on real tasks.
  • Check the context length actually configured, because many deployments use less than the model supports.
  • Add guard models or filters when exposing Llama to the public, since self-hosting removes provider-side safeguards.

In simple terms

Llama is a family of AI language models made by Meta. Most chatbot models can only be used through a company’s website. Llama is different: you can download the model itself and run it on your own computers.

It is a bit like the difference between eating at a restaurant and getting the recipe. With the recipe, you can cook at home, change the ingredients and keep your kitchen private. The trade-off is that you do the cooking: you need suitable hardware, some technical skill and your own safety checks.

A practical example

A mid-sized law firm wants an assistant that summarizes contracts, but client confidentiality rules forbid sending documents to an external AI service. The firm’s IT team deploys a Llama 70B instruct model on two in-house GPU servers using an open-source inference server, with the model quantized to 8 bits to fit the available memory.

The team then applies a LoRA fine-tune on about 3,000 internal summaries written by senior associates, teaching the model the firm’s preferred structure and terminology. Before rollout, lawyers compare 150 model summaries with human ones and flag errors in clause interpretation, which leads the firm to require human review for every summary sent to a client.

The result is a private drafting assistant that keeps all documents on-premises, cuts first-draft time for routine contracts by several hours a week per lawyer, and runs under the license terms the firm’s counsel has reviewed.