Llama

Llama is a family of language models released by Meta, beginning with a research model in February 2023 and continuing through several generations of increasingly capable open-weight systems. Unlike the flagship models of OpenAI, Anthropic and Google, Llama's trained weights are published for download, so anyone with the hardware can run, modify or fine-tune them without permission from Meta 1.

That decision made Llama the reference point for open-weight AI. A large part of the research community, and most start-ups that could not afford to train a model, built on Llama or its derivatives. At the same time, the licence is not an open-source licence in the recognised sense, and Meta has restricted some uses and some regions, which has made the meaning of "open" a continuing argument 2.

This entry covers the model generations, the architecture and training that Meta has disclosed, the benchmarks and their caveats, the licensing model, and the disputes — over benchmark reporting, regional availability and the safety of releasing weights — that have followed the family.

The distinction matters. A model's weights are the numbers produced by training; publishing them allows anyone to run inference, fine-tune and redistribute the model. Open-source software, as the term is normally used, also grants the freedom to use the work for any purpose and to study and modify it without restriction. Meta's Llama licence grants most of those freedoms but adds conditions, so the Open Source Initiative does not recognise it as open source 2.

The conditions are not trivial. The licence includes an acceptable-use policy, requires attribution and naming conventions for derivatives, and contains a clause that applies to products with more than 700 million monthly active users, which requires a separate arrangement with Meta. It also excludes some jurisdictions and some multimodal versions from release in the European Union 2.

In practice the licence has been permissive enough for most users. Cloud providers host Llama models, companies fine-tune them on private data, and researchers publish derivatives. The restrictions matter most at the extremes: a very large platform that wants to build on Llama, or a user in a jurisdiction Meta has excluded, faces limits that a genuinely open licence would not impose 3.

The label also carries commercial value. Calling Llama open source signals community friendliness and invites contributions, while retaining control over the largest deployments and over the brand. Meta's critics describe this as open-washing; Meta describes it as a pragmatic balance between openness and responsible use 4.

LLaMA was published in February 2023 in sizes from 7 billion to 65 billion parameters, trained on between one and 1.4 trillion tokens and intended for research. Within weeks the weights had been posted online without authorisation, and the model spread through the research community far faster than Meta had planned. The leak made LLaMA the foundation of a wave of open models 1.

Llama 2 followed in July 2023 with 7, 13 and 70 billion parameter versions, a commercial licence, and chat models fine-tuned with reinforcement learning from human feedback. It was trained on two trillion tokens and shipped with a detailed safety report, including red-teaming results. Llama 2 established the pattern that later releases followed: a base model, a chat model and a licence 5.

Llama 3 arrived in April 2024 with 8 and 70 billion parameter models trained on more than fifteen trillion tokens, a larger vocabulary and better multilingual performance. Llama 3.1 in July 2024 added a 405 billion parameter model, a 128,000-token context window and tool calling. Llama 3.2 followed in September with small text models and vision-language versions, and Llama 3.3 in December delivered performance close to the 405 billion parameter model from a 70 billion parameter model that could run on a single high-end server 6 7.

Llama 4, released in April 2025, changed the architecture. Scout and Maverick are mixture-of-experts models, activating a fraction of their parameters per token, and were trained natively on multimodal data. A larger model, Behemoth, was described as still in training and was not released. The generation was positioned as Meta's answer to frontier proprietary systems, though its reception was cooler than earlier releases 8.

The first three generations are decoder-only transformers of a conventional design: pre-normalisation, rotary position embeddings, grouped-query attention in later versions, and a SwiGLU feed-forward block. Meta published these details in technical reports, which is unusual for a frontier lab and one reason Llama became a standard research baseline 6.

Llama 4 introduced mixture-of-experts layers. Scout has 109 billion total parameters and activates 17 billion per token; Maverick has 400 billion total and activates 17 billion; Behemoth, the unreleased model, was described as having close to two trillion. The design lowers the compute cost of inference relative to a dense model of the same total size, at the cost of memory to hold all the experts 8.

Training data is not published as a dataset, but Meta has described its composition in broad terms: web text, code, multilingual material and, for Llama 4, images and video. Llama 3 was trained on more than fifteen trillion tokens, and Meta has said it invested heavily in data filtering and de-duplication, including classifiers that remove low-quality pages 6.

The safety approach has shifted across generations. Llama 2 was fine-tuned with supervised safety data and reinforcement learning from human feedback, and Meta published red-team results. Later models added more extensive evaluations and a system card, but the fundamental problem of open weights remains: once a model is downloaded, the developer cannot enforce its safety policy, and fine-tuning can remove refusal behaviour 5.

Llama models are general-purpose text generators. The larger versions handle long-document summarisation, code generation, translation and instruction following competently, and fine-tunes can be specialised for tasks from medical coding to legal drafting. The smaller versions run on a laptop or a single GPU, which makes them useful for privacy-sensitive work where data cannot leave a device 7.

The models are not at the frontier. On hard reasoning, mathematics and long-horizon coding they trail the best proprietary systems, and the gap widened after 2025 as competitors released stronger reasoning models. Llama's advantage is control and cost, not raw capability, and Meta's own assistant uses a mixture of models rather than Llama alone 8.

Multimodality arrived with Llama 3.2's vision models and became native in Llama 4. Image understanding works well enough for document analysis and scene description; video is handled as a sequence of frames and remains limited. The models do not generate images or audio 8.

Practical limits include memory, because a 405 billion parameter model needs multiple high-end accelerators even in reduced precision, and context, because long windows degrade in effective use. Fine-tuning is cheap relative to pre-training but still requires expertise, and a badly tuned model can lose the safety behaviour that Meta built in 3.

Meta reports benchmark results with each release, and the numbers are vendor-reported. Llama 3.1 405B was reported at 88.6 per cent on MMLU, 96.8 per cent on GSM8K and 89.0 per cent on HumanEval, roughly competitive with the best proprietary models of the time. Llama 4 Maverick was reported at 80.5 per cent on MMLU Pro and strong results on multimodal benchmarks, while Scout was positioned as cheaper and longer-context 6 8.

These figures require care. MMLU is saturated, HumanEval is small and may be contaminated, and mixture-of-experts models can behave differently in a harness than their benchmark numbers suggest. Meta has published harness details for some evaluations, but independent replication is partial and often finds smaller margins than the release post claims 4.

The most serious dispute concerned Llama 4 and an arena leaderboard. Meta submitted an experimental chat version of Maverick to the LMArena comparison site, where it ranked near the top, and then released a different model to the public. Researchers noted the discrepancy; Meta acknowledged that the tested version was tuned for conversational preference and that the released model differed. LMArena subsequently revised its policies to require disclosure when a tested model is not the released one 4.

For downstream users the practical lesson is that a benchmark score describes a checkpoint, not a family. A fine-tune, a quantised version or a different system prompt can move performance substantially, and the only meaningful evaluation is one run on the task the model will actually perform 3.

Llama weights are distributed through Meta's own site and through Hugging Face, and downloads are gated: users accept the licence and provide basic details before access is granted. The models are also available through cloud platforms including AWS, Azure, Google Cloud and specialised inference providers, which charge by the token and handle the hardware. For most companies this hosted route is cheaper than running the model themselves 3.

The licence permits commercial use, modification and redistribution, with conditions on naming, acceptable use and scale. It does not grant patent rights in the way some open-source licences do, and it can be revised by Meta for later versions, which means that a derivative may be governed by the terms attached to the version it came from rather than the current ones 2.

Regional availability has been uneven. Meta initially withheld the multimodal Llama 3.2 models from the European Union, citing regulatory uncertainty, and Llama 4's European rollout was delayed. Users in those markets could sometimes obtain the models through third parties, which undermined the restriction and created confusion about what was permitted 4.

Meta does not charge for the weights, so its return comes indirectly: through advertising and engagement in its products, through goodwill with developers, and through the strategic benefit of ensuring that a capable model is widely available rather than concentrated in competitors' hands. That logic explains both the openness and its limits 2.

The openness debate is the central criticism. Supporters argue that downloadable weights democratise access, allow scrutiny, prevent lock-in and support research that would otherwise be impossible outside a few companies. Critics argue that open weights remove the ability to recall a model, that fine-tuning strips safety measures, and that the same weights can be used for disinformation, malware and non-consensual imagery. Both sides agree on the facts and disagree about how to weigh them 4.

The Llama 4 launch damaged Meta's credibility on measurement. The arena episode suggested that the company had optimised a model for a public leaderboard rather than for the product it shipped, and the gap between the launch claims and independent results was widely noted. Meta's explanation — that the tested model was an experimental variant — was accurate and unsatisfying, and it fed a broader scepticism about vendor benchmarks 4.

The licence has attracted criticism from the open-source community. The 700 million user threshold, the acceptable-use policy and the regional exclusions are incompatible with the Open Source Initiative's definition, and calling Llama open source has been described as misleading. Meta has not claimed OSI certification, but its marketing has not always been careful about the distinction 2.

Internally, the Llama effort has been affected by the reorganisation of Meta's AI divisions, including layoffs in 2025 that reduced research staff and shifted resources toward a new superintelligence unit. The effect on future releases is uncertain, but the departure of senior researchers and the delay of Behemoth suggested that the open-weight programme was no longer the company's only priority 8.

Whether Llama is genuinely competitive is contested. Meta's benchmarks place it close to the frontier at each release, and independent evaluations often place it a little below. The gap is explained partly by harness differences and partly by the fact that Meta benchmarks base models while users interact with fine-tuned chat versions. The fair conclusion is that Llama is competitive for an open model and generally behind the best closed ones 6.

The safety argument is more fundamental. Meta's position is that open weights are safer because they allow external scrutiny and because the alternative concentrates power in a few laboratories. Critics reply that scrutiny has not prevented misuse, that the company does not track downloads meaningfully, and that a model which cannot be withdrawn is a different kind of risk from one that can. Neither claim has been tested at the scale that would settle it 5.

The "open source" label is disputed in a way that is easy to resolve in principle and hard in practice. By the OSI definition, Llama is not open source; by the looser definition used in much of the AI industry, it is. The ambiguity benefits Meta, because it gains the reputation of openness without the obligations, and it frustrates those who want the term to mean something precise 2.

What is not in doubt is the effect on the field. Before Llama, a capable language model was something a company accessed through an API. After it, a capable model was something a university group could download, inspect and modify. That shift is the family's main historical contribution, and it happened partly against Meta's intention when the first weights leaked 1.