In February 2023, a language model called LLaMA appeared on 4chan as a BitTorrent link. Meta had intended to hand it out to approved researchers only. Within a week the weights were running on laptops and graduate students’ desktops, and the company’s careful distribution plan was gone. That leak set the tone for the next two years: Meta kept publishing, and the world kept downloading.
Today Meta Llama is the most widely used family of open-weight language models on the planet. It runs inside WhatsApp, sits on AWS and Azure, gets fine-tuned by hospitals and law firms, and powers a long tail of startups that could never afford to train a model themselves. Here’s what it actually is, what the license really says, and where it makes sense to use it.
What Meta Llama actually is
Llama is a series of transformer-based language models built by Meta’s AI research teams and released as downloadable weights. You don’t have to call an API to use them. You download a file measured in gigabytes, point an inference engine at it, and the model runs on hardware you control.
The models come in sizes from 1 billion to 405 billion parameters, plus a couple of Mixture-of-Experts variants in the Llama 4 generation that activate only a slice of their total parameters per token. That range matters more than any single benchmark number, because it means the same architecture family can run on a Raspberry Pi and on a data centre rack.
Open weights is not the same as open source
This distinction trips people up constantly. Llama’s weights are downloadable, but Meta does not publish the training data, the full data pipeline, or the code used to build the corpus. The license is a custom community agreement rather than an OSI-approved open source license, and it carries conditions. So Llama is “open weight” in the way a recipe handed out at a restaurant is open: you can cook it, tweak it, sell the dish, but you can’t claim you invented it and you’re bound by a few house rules.
A quick tour of the family
The version history is worth knowing because deployment guides and fine-tuning tutorials online often assume a specific generation:
- Llama 2 (July 2023) — the first commercially licensed release, 7B to 70B. Still floating around in legacy deployments.
- Llama 3 (April 2024) — 8B and 70B, trained on roughly 15 trillion tokens, a large jump in quality.
- Llama 3.1 (July 2024) — added the 405B flagship and a 128K context window across the range.
- Llama 3.2 (September 2024) — tiny 1B and 3B models for phones and edge devices, plus 11B and 90B vision-language variants.
- Llama 3.3 70B (December 2024) — a 70B model tuned to roughly match the 405B on many tasks at a fraction of the serving cost.
- Llama 4 Scout and Maverick (April 2025) — mixture-of-experts models with native multimodality; Scout advertises a 10-million-token context window.
If you’re starting a project today, Llama 4 Maverick or a 3.3 70B checkpoint is usually the sensible default, with the 3.2 small models reserved for on-device work.
Why giving the weights away was a smart move
Meta doesn’t sell model access. Its business is advertising, and advertising benefits when more software exists, when developers default to your ecosystem, and when nobody can charge you rent for the underlying intelligence.
Releasing Llama did three things at once. It commoditised the model layer, which squeezed the margins of anyone trying to resell inference. It made Meta the reference point for fine-tuning, so an entire tooling industry grew up around its chat templates and tokenizers. And it created goodwill among developers who might otherwise have standardised on a competitor. If you want the wider picture of what this open-weight approach enables in practice, there’s a good breakdown of what you can actually run, own and ship with open source AI.
You’ve almost certainly used it already
Meta AI, the assistant baked into WhatsApp, Instagram, Messenger and Facebook, runs on Llama. It sits in a search bar in an app you opened this morning. The company has pushed it aggressively, and the rollout has been so broad that the assistant now appears in places people didn’t ask for it. That dynamic is worth understanding, and this piece on Meta AI showing up everywhere you look covers the strategy behind it.
Beyond Meta’s own products, Llama weights are hosted by AWS Bedrock, Azure AI, Google Cloud Vertex, Together, Fireworks, Groq and dozens of smaller providers. A startup building a support chatbot on Bedrock in 2024 was, in all likelihood, running Llama without ever mentioning it in a pitch deck.
How it stacks up against the closed models
Honest answer: the top closed models from OpenAI, Anthropic and Google still lead on the hardest reasoning, long-horizon agentic tasks and messy real-world instruction following. Llama’s advantage isn’t raw benchmark supremacy. It’s cost, control and the ability to change the model’s behaviour without asking permission.
At small and mid sizes the gap is narrower than the leaderboards suggest, especially after task-specific fine-tuning. If you’re weighing options across vendors, this comparison of ChatGPT alternatives picked by what you actually need is a more useful frame than any single score.
What running Llama yourself actually costs
Rough numbers, assuming 4-bit quantisation:
- 1B–3B: fits in a few gigabytes of RAM. Runs on a phone, a Raspberry Pi 5, or a browser tab.
- 8B: around 5–6 GB. Comfortable on a 16 GB laptop with a decent GPU.
- 70B: around 40 GB. Needs something like two 24 GB cards, a Mac Studio with 64 GB of unified memory, or rented cloud capacity.
- 405B: north of 200 GB even quantised. This is a data centre model, not a desktop one.
Tooling is not the bottleneck anymore. Ollama gets you a working endpoint in about two commands. llama.cpp handles CPU and Apple Silicon well. vLLM is what you reach for when you need to serve many concurrent users, because it batches requests efficiently instead of processing them one at a time.
When latency is the whole product
Serving speed is its own axis, separate from model quality. Cerebras offers Llama inference at token rates that feel less like streaming and more like pasting, thanks to wafer-scale chips that keep the whole model on-die. The engineering behind that is genuinely unusual, and it’s covered well in Cerebras Systems refusing to cut the wafer.
The license fine print worth reading
The Llama 4 community license permits commercial use, including fine-tuning and distillation, with one clause that matters: companies with more than 700 million monthly active users need a separate agreement from Meta. That threshold is aimed squarely at the handful of giants who could challenge Meta directly. For everyone else it’s effectively permissive.
Two other conditions are easy to miss. You must display “Llama” prominently in the model name of anything you distribute, and derivatives must carry the prefix “Llama” at the start of the name. Fine-tune Llama 4 Scout into a legal research model and it ships as something like “Llama-Lex-Scout”, not “LexScout”. Also, the acceptable use policy applies downstream, so your users are bound by Meta’s rules too.
Where Llama genuinely earns its place
Four situations where the open weights pay off in ways an API can’t:
Data you can’t send anywhere. A hospital summarising patient notes or a bank triaging internal tickets has a hard boundary around data leaving the building. Self-hosted inference ends that conversation.
Volume economics. Past a certain request threshold, a reserved GPU costs less than per-token pricing, and the break-even point arrives faster than most teams expect.
Behaviour you need to pin down. Fine-tuning on 5,000 of your own examples routinely beats prompt engineering on a 500-billion-parameter generalist for narrow, repetitive tasks. You also freeze the model version, so your product doesn’t quietly change behaviour when a vendor ships an update.
Languages and domains the big labs under-serve. A community of Shona or Bengali-language fine-tunes exists precisely because someone could download the weights and do the work.
Pick against the smallest model that clears your quality bar, not the largest one you can afford. A 3B fine-tune that answers your specific question correctly beats a 405B generalist that sometimes wanders. Test on your own data, watch the cost per thousand requests, and keep an eye on the next release, because Meta has shipped a new Llama generation roughly every six to nine months and shows no sign of slowing.