Llama: Meta's Open-Weight Model Family

What Llama Is

Llama (originally stylized LLaMA, for Large Language Model Meta AI) is a family of large language models developed by Meta Platforms and released with downloadable model weights. Unlike closed frontier systems that are reachable only through a hosted API, every published Llama model ships its trained parameters so that developers can run, inspect, fine-tune, and deploy the model on their own hardware.

Meta positions Llama as the centerpiece of an openly available AI ecosystem. The design philosophy favors broad accessibility and a large downstream developer community over keeping the weights proprietary. As of 2026 the lineage runs from the research-only Llama 1 (February 2023) through the commercially licensed Llama 2 and Llama 3 generations to the Llama 4 herd, which Meta describes as its first natively multimodal, mixture-of-experts family.

An important caveat applies to the whole family: Llama is open-weight, not open-source in the strict sense. The weights are downloadable, but distribution is governed by a custom Meta community license with usage restrictions rather than an OSI-approved open-source license.

• Maker: Meta Platforms (formerly Facebook).
• Distribution: downloadable weights, not API-only.
• Status: open-weight under a custom license, not formally open-source.
• Scope as of 2026: Llama 1 through the Llama 4 herd.

Open-Weight vs Open-Source: What the Llama License Allows

The terms open-weight and open-source are often conflated but mean different things. Open-source, in the formal sense used by the Open Source Initiative, requires that anyone may use the software for any purpose, with no restriction on fields of endeavor or categories of user. Open-weight is narrower: the trained parameters are downloadable, but the accompanying license can impose conditions on who may use them and how.

The Llama Community License is a custom Meta agreement, not a standard open-source license. It grants a broad, royalty-free right to use, reproduce, distribute, and create derivative works from the weights, but it attaches conditions that an OSI-approved license would not permit. For this reason commentators classify Llama as open-weight rather than open-source.

• Open-weight: parameters are downloadable and runnable offline, but the license may restrict certain users or uses.
• Open-source (OSI definition): no restrictions on field of use or class of user; Llama does not meet this bar.
• Practical effect: most developers and companies can use Llama freely, but a few categories (very large platforms and certain EU uses) are constrained.

The Llama Lineage: From Llama 1 and 2 Through Llama 3 to Llama 4

The family has evolved rapidly. The figures below are accurate as of 2026 and reflect Meta's public releases.

Each generation broadened access and capability: Llama 1 was research-only, Llama 2 introduced commercial use, Llama 3 added much stronger reasoning and the very large 405B variant, and Llama 4 shifted the architecture to mixture-of-experts with native multimodality.

• Llama 1 (February 2023): 7B to 65B parameters, released under a non-commercial research license.
• Llama 2 (July 2023): 7B, 13B, and 70B parameters, the first generation permitted for commercial use under the community license.
• Llama 3 (April 18, 2024): 8B and 70B dense models.
• Llama 3.1 (July 23, 2024): added the 405B parameter flagship alongside refreshed 8B and 70B models, with longer context and multilingual support.
• Llama 3.2 (late 2024): added small on-device 1B and 3B text models plus the family's first vision (multimodal) models.
• Llama 4 (April 5, 2025): the mixture-of-experts, natively multimodal herd (Scout and Maverick shipped; Behemoth previewed).

The Llama 4 Herd: Scout, Maverick, and Behemoth

Released on April 5, 2025, the Llama 4 herd is Meta's first family built on a mixture-of-experts (MoE) design, in which only a subset of the model's parameters (the active parameters) is engaged for any given token. This lets a model hold a very large total parameter count while keeping inference cost closer to a much smaller dense model. The specifications below are as of 2026.

Scout and Maverick were released as open weights under the Llama 4 Community License. Behemoth was announced as a preview and, as of 2026, has not been released as open weights; reporting through 2025 and into 2026 indicates Meta delayed or paused its launch over capability concerns. Behemoth functions primarily as a teacher model used to distill knowledge into the smaller variants.

• Llama 4 Scout: 17B active parameters, 16 experts, roughly 109B total parameters, with an advertised context window of up to 10 million tokens.
• Llama 4 Maverick: 17B active parameters, 128 experts, roughly 400B total parameters, with a context window of up to 1 million tokens.
• Llama 4 Behemoth (previewed, unreleased as of 2026): about 288B active parameters, 16 experts, and on the order of 2 trillion total parameters; used internally as a teacher model.
• All three are natively multimodal, trained jointly on text and images rather than bolting vision on afterward.

Mixture-of-Experts and Native Multimodality

Mixture-of-experts replaces some dense feed-forward layers with a set of parallel expert subnetworks plus a router that selects a small number of experts per token. The result is a model with a large total capacity but a much smaller active footprint at inference time. In Llama 4, for example, Maverick carries roughly 400B total parameters but activates only about 17B per token, which improves throughput and cost relative to a dense model of comparable total size.

Native multimodality means the Llama 4 models were pre-trained from the start on a mixed corpus of text and images (Meta reported a training mixture exceeding 30 trillion tokens spanning text, image, and video data). This contrasts with earlier approaches that attached a separate vision encoder to a text-only base after the fact. The combination of MoE efficiency and early-fusion multimodality is the defining technical shift of the Llama 4 generation.

• MoE: large total capacity, small active compute; only a few experts run per token.
• Native (early-fusion) multimodality: text and image understanding trained together from the start.
• Long context: Scout's advertised 10M-token window targets large-document and multi-file workloads, though real-world quality at extreme lengths depends on the task.

Licensing and Restrictions as of 2026

Two restrictions in the Llama Community License are most consequential for adopters, and both remain in force as of 2026.

First, the very large platform clause: if a product or service (counting the licensee and its affiliates) exceeded 700 million monthly active users in the month before the relevant Llama version was released, that company must request a separate license from Meta, which Meta may grant or deny at its sole discretion. This effectively carves out the largest technology companies from the default grant.

Second, an EU-specific limitation introduced with Llama 3.2 and carried forward: for multimodal Llama models, the license grant is not extended to individuals domiciled in, or companies with a principal place of business in, the European Union. End users of a product that merely incorporates such a model are exempted, and non-EU companies may still distribute products containing the models under their normal global practices. Adopters should always read the specific version's license and Acceptable Use Policy, since terms have changed between releases.

• 700M MAU clause: the largest platforms must seek a separate license from Meta.
• EU multimodal limitation (since Llama 3.2): the grant excludes EU-domiciled individuals and EU-headquartered companies for multimodal models, with carve-outs for end users.
• Acceptable Use Policy: prohibits illegal and harmful uses and is incorporated into the agreement by reference.
• Terms vary by version; verify the exact license for the specific Llama release you deploy.

Why Open Weights Matter: Fine-Tuning, Self-Hosting, and Privacy

Downloadable weights change the economics and control profile of using a model. Because the parameters run locally, the same Llama checkpoint can be fine-tuned on domain data, quantized to fit consumer GPUs, served behind a private network, or embedded in an offline application, all without sending data to a third-party API.

For privacy-sensitive and regulated workloads this matters: inputs never have to leave controlled infrastructure, and the model's behavior is reproducible and auditable. This is one reason open-weight models are popular for on-device assistants and personal knowledge tools. AI memory and second-brain applications such as MemX, which retrieve a user's documents, photos, and voice notes by plain-English queries, are a natural fit for locally hostable models when keeping personal data on-device is a priority.

• Fine-tuning and adapters (for example LoRA) let teams specialize a base model cheaply.
• Self-hosting removes per-token API costs and external dependencies.
• Local inference keeps sensitive data inside controlled boundaries.
• Quantization (8-bit, 4-bit and lower) shrinks models to run on laptops and single GPUs.

Llama in the Ecosystem: Ollama, llama.cpp, and Local Second Brains

A large open-source tooling ecosystem has grown around Llama. llama.cpp is a C/C++ inference engine that runs quantized Llama-family models efficiently on CPUs and consumer GPUs and popularized the GGUF model format. Ollama provides a simple command-line and API wrapper for pulling and running models locally, building on similar inference techniques. Hugging Face hosts the official weights and a large catalog of community fine-tunes.

These tools make it practical to run a capable model on a personal machine, which underpins local second-brain and retrieval-augmented setups. A typical pattern pairs a local Llama model with a vector database and semantic search so that a user's notes and files can be queried in natural language without cloud inference.

• llama.cpp: efficient local inference and the GGUF quantized format.
• Ollama: easy local model management and a built-in serving API.
• Hugging Face: official and community Llama weights and fine-tunes.
• Common stack: local Llama model plus embeddings and a vector store for private retrieval.

Limitations and Comparison to Closed Frontier Models

Open weights bring real trade-offs. On the hardest reasoning, coding, and agentic benchmarks, the largest closed frontier models from the leading labs have generally held a lead over openly available models, and that comparison shifts with every release, so any specific ranking is volatile and should be checked against current evaluations.

Running large Llama models also demands significant hardware: the biggest variants need multi-GPU or multi-node infrastructure, putting them out of reach for casual local use even though smaller and quantized variants run on a single machine. Licensing is not unrestricted, as noted above. And Behemoth, the most capable Llama 4 tier, remained unreleased as open weights as of 2026, so the openly available ceiling is set by Scout and Maverick. For many practical applications, however, the combination of strong mid-size models, full control, and no per-token cost makes open-weight Llama a compelling alternative to closed APIs.

• Closed frontier models often lead on the hardest benchmarks, but rankings change frequently and should be re-verified.
• The largest Llama variants require substantial GPU infrastructure to run.
• Licensing is restricted, not fully open; the 700M MAU and EU multimodal clauses apply.
• The strongest tier (Behemoth) was not released as open weights as of 2026.