Llama

There is no Llama app to sign up for and no Llama Pro tier. Meta's consumer assistant is a separate product that lives in the Meta AI app and inside Instagram, WhatsApp and Facebook. What "Llama" names is a set of model files that Meta publishes for download, together with the licence that governs what you may do with them. You either run those files on hardware you control, or you pay somebody else who has already loaded them onto theirs. That single distinction drives every practical decision on this page.

That structure is why the family became the default open-weights choice. A downloadable model can be fine-tuned on proprietary data, quantised to fit a specific GPU, pinned to a version that never changes underneath you, and run inside a network that has no route to the public internet. None of that is possible with a hosted frontier API, however good it is.

As of August 2026 the family Meta publishes looks like this:

• Llama 4 Scout — 17B active parameters of 109B total across 16 experts, natively multimodal (text and image in, text and code out), with a documented 10-million-token context window and pretraining on roughly 40 trillion tokens.
• Llama 4 Maverick — 17B active of roughly 400B total across 128 experts, same modalities, 1-million-token context, pretrained on roughly 22 trillion tokens. An FP8 build is published alongside.
• Llama 3.3 70B Instruct — the newest dense, text-only instruct model, from December 2024.
• Llama 3.2 1B and 3B — small multilingual text models for on-device work, plus 3.2 Vision at 11B and 90B.
• Llama Guard 4 12B and Prompt Guard 2 — open safety classifiers for screening inputs and outputs, and for flagging jailbreak and prompt-injection attempts before a request reaches your main model.

Both Llama 4 models officially support twelve languages — Arabic, English, French, German, Hindi, Indonesian, Italian, Portuguese, Spanish, Tagalog, Thai and Vietnamese — and carry an August 2024 knowledge cutoff. Meta's model card is explicit that pretraining data included publicly shared posts from Instagram and Facebook alongside licensed and public sources.

The most important fact about Llama in 2026 is not in the model cards. Meta has not shipped a new open-weights generation since April 2025. Its 2026 flagship work comes from Meta Superintelligence Labs under an entirely different brand: Muse Spark launched in April 2026, Muse Spark 1.1 followed on 9 July 2026 as a multimodal reasoning model aimed at agentic tool use and coding with a 1-million-token context, and Muse Image and Muse Video arrived the same month. Those models are reached through the new Meta Model API, in public preview, and through the Meta AI app — not through a download link. Searching Hugging Face for Muse weights returns nothing.

Meanwhile llama.com itself now redirects to developer.meta.com. Read that as a signal rather than a death notice: Llama 4 is still served, still supported by every major cloud, and Meta still publishes open models in vision and science. But anyone planning a multi-year build on the assumption that a Llama 5 will arrive to replace Llama 4 is planning on something Meta has not announced.

The Llama 4 Community Licence is permissive but it is not Apache or MIT, and the differences matter if you are shipping a product.

You get a non-exclusive, worldwide, royalty-free licence to use, reproduce, distribute, modify and create derivative works. In exchange, redistribution obliges you to include the agreement itself, retain an attribution notice in a NOTICE file, and prominently display "Built with Llama" on a related website, interface, blog post or product documentation. If you use Llama or its outputs to train or fine-tune another model that you distribute, that model's name must begin with "Llama" — a real constraint if you intended to brand your fine-tune as something else entirely. Use also has to comply with the Acceptable Use Policy, which is incorporated into the licence by reference.

The clause everyone quotes is Section 2: if the products you or your affiliates offer had more than 700 million monthly active users in the calendar month before the version's release date, you must ask Meta for a separate licence, and Meta may refuse. In practice this is a clause aimed at a handful of companies. For everyone else it is noise. Two smaller details are worth knowing: you own the derivative works and modifications you make, and the licence terminates if you sue Meta over Llama-related IP. Governing law is California.

Licence fee zero, hardware bill not zero. The rough arithmetic follows from parameter counts and precision. A 1B or 3B Llama 3.2 model quantised to 4-bit fits comfortably on a laptop. Llama 3.1 8B needs single-digit gigabytes and runs on one consumer GPU. Llama 3.3 70B at 4-bit lands near 40GB, which realistically means one 48GB card or two 24GB cards, plus enough headroom for the KV cache once contexts get long.

Llama 4 breaks the intuition that mixture-of-experts models are cheap to host. Only 17B parameters activate per token, which makes them fast and cheap to run — but all the experts still have to sit in memory. Scout's 109B total needs roughly 60GB even at 4-bit; Maverick's ~400B is a multi-GPU server, and the published FP8 build exists precisely because the full-precision one is impractical for most people. MoE saves you compute, not memory.

Then add the parts that do not appear in a spec sheet: batching and serving infrastructure, evaluation harnesses when you upgrade, someone on call when throughput collapses at 2am, and the quantisation work needed to make any of it fit your budget.

Llama is the most widely hosted open family in the industry, which means self-hosting is a choice rather than a requirement. AWS Bedrock lists Llama 4 Scout and Maverick alongside the 3.x models. Microsoft Foundry, the renamed Azure model catalogue, carries both Llama 4 models. Google serves Scout through Vertex, and specialist inference providers including Groq, DeepInfra and Novita compete on the same weights — visible on aggregators such as OpenRouter, where Scout was listed around $0.10 per million input and $0.30 per million output tokens in August 2026, and Maverick around $0.20 and $0.70.

Those numbers reframe the whole question. At that rate, a workload has to be substantial before a GPU lease plus an engineer's time beats simply paying per token. The good news is that the portability is real: because every host serves the same weights, moving between them is a configuration change rather than a rewrite and re-evaluation.

Self-host when the reason is structural rather than financial — data that legally cannot leave your estate, an air-gapped deployment, a fine-tune on proprietary data that you need to keep private, or a product that must guarantee the same model behaviour for years. Those are things no API can sell you.

Rent when the reason is capability or convenience. If you want the best available reasoning, agentic coding or up-to-date knowledge, the frontier is elsewhere, and Meta's own newest models are no longer part of the open story. Between those poles, the pragmatic pattern is to call Llama through a hosted endpoint while you prototype, measure whether a smaller fine-tuned model would do the job, and only take on GPUs once the volume or the compliance requirement makes that unavoidable.