TL;DR: Llama is Meta's open-weight model family. You download the weights and run them yourself, or use one of many cheap hosted providers. Llama 4 Scout has a huge 10M token context window. Maverick is the 400B flagship. Behemoth, near 2 trillion parameters, is still in training. Hosted prices start near $0.08 per million tokens.
Llama is the most widely deployed open model family in the West. Every Llama 4 model is open-weight, so you can download it from Hugging Face and run it on your own hardware, or rent it from a hosting provider. The current herd has three members. Llama 4 Scout is small, fast, and carries a record 10M token context window. Llama 4 Maverick is the 400B flagship for chat and coding. Llama 4 Behemoth, near 2 trillion parameters, is still training and not yet public.
Key takeaways
- Llama 4 is open-weight and natively multimodal, built on a Mixture-of-Experts design.
- Scout has a 10M token context window, the largest in any shipping open model.
- Maverick is the 400B flagship. Behemoth, near 2T parameters, is still in training.
- Run it two ways: self-host the free weights, or use a cheap host like Together, Groq, or Bedrock.
- The Llama license allows commercial use, except for companies with over 700M monthly users.
The Llama models at a glance
Here is the lineup, with parameters and context. MoE models list active and total parameters, because only the active slice runs on each token. Hosted prices are typical blended rates from providers, not a single official price. Use this table as your reference.
| Model | Parameters | Context | Good for |
|---|---|---|---|
| Llama 4 Scout | 17B active / 109B (16 experts) | 10M | Long context, fits one GPU |
| Llama 4 Maverick | 17B active / 400B (128 experts) | 1M | Flagship multimodal, chat and code |
| Llama 4 Behemoth | 288B active / ~2T | — | Teacher model, still training |
| Llama 3.3 70B | 70B dense | 128K | Proven dense workhorse |
| Llama 3.1 8B | 8B dense | 128K | Edge, fine-tuning, cheap volume |
How do you want to run Llama?
Pick your setup to see the right model and the team you need.
For a fast start with no infra, run Llama 4 on Together, Groq, Fireworks, or Bedrock. Blended rates start near $0.08 per million tokens. See what building this in-house costs on our developer rate card.
For full data control and zero per-token cost, run Scout or Maverick on your own GPUs. This needs strong deployment skills. Hire an AI model deployment engineer who runs vLLM and SGLang in production.
For very long documents or image tasks, Scout reads up to 10M tokens and both models are natively multimodal. Strong AI and machine learning engineers get the most from them.
The model is the easy part. The hard part is the team that deploys and ships with it. We match you with pre-vetted engineers in about 24 hours. Tell us what you need.
The Llama 4 herd in plain English
Llama 4 is the first Llama built on a Mixture-of-Experts design and the first that is natively multimodal, meaning it reads text and images in one model. An MoE model only activates a small slice of its parameters on each token, so it runs far cheaper than its total size suggests. That is how a 400B model like Maverick serves at the cost of a much smaller one.
Scout, Maverick, and Behemoth
Scout is the small, fast member. It uses 17B active out of 109B total parameters, fits on a single high-end GPU, and reads up to 10M tokens, the largest context window in any shipping open model. Maverick is the flagship, with 17B active out of 400B total across 128 experts, tuned for chat and coding at a 1M context window. Behemoth is the giant teacher model, near 2 trillion parameters, still in training and not released as of mid-2026.
The Llama 3 models
The older Llama 3 line is still widely used. These are dense models, meaning all parameters run on every token. Llama 3.3 70B is a proven workhorse for chat and reasoning at a 128K context window. Llama 3.1 comes in 8B, 70B, and 405B sizes. The small 8B model is the one most teams fine-tune, because it is cheap to train and easy to deploy on modest hardware.
Open weights and the license
Llama is open-weight, but the license is not the fully permissive Apache 2.0 that some rivals use. The Llama Community License lets almost any company use the models commercially, for free. The main limit is scale: a company with over 700 million monthly active users must request a separate license from Meta. For the vast majority of teams, that limit never applies.
Because the weights are open, you can run Llama anywhere. Self-host it on your own GPUs for full control and zero per-token cost. Or rent it from a hosted provider such as Together, Groq, Fireworks, DeepInfra, or Amazon Bedrock, and pay per token with no infra to manage. Running any of these well is a job for strong AI and machine learning engineers.
Pricing and where to run it
There is no single Llama price, because Meta does not run a first-party API. Instead many providers host the models and compete on rate. Blended prices run from about $0.08 per million tokens for the small models up to a few dollars for the largest. That is often 5 to 10 times cheaper than a frontier closed model for the same task. The chart shows typical hosted output prices.
Self-hosting changes the math. You pay $0 per token and instead pay for GPUs and the engineers who run them. For steady, high-volume work, that fixed cost often beats per-token pricing by a wide margin. For spiky or low-volume work, a hosted provider usually wins. The right answer depends on your traffic, and getting it right is a real engineering decision.
Context windows and multimodal input
Scout’s 10M token window is the headline. That is roughly 7.5 million words, enough to load an entire large codebase or a shelf of documents into one request. Maverick offers a 1M window, and the Llama 3 models 128K. In practice, very long context needs a lot of memory to serve, so real deployments often cap it below the maximum to control cost.
Because Llama 4 is natively multimodal, the same model reads images and text together. That helps with screenshots, charts, and document understanding without a separate vision model. An engineer who understands MoE serving, context limits, and multimodal input will run Llama far cheaper at the same quality. You can read more in our breakdown of Claude versus ChatGPT for coding.
Which model should you use
Start by choosing the path. If you need data control or run very high volume, self-host. If you want to ship fast with no infra, use a hosted provider. From there, pick Maverick as the default, Scout for very long context, and a Llama 3 model for cheap fine-tuning. The table below maps common jobs to a model.
| Job | Recommended model | Why |
|---|---|---|
| Support chatbot, FAQ | Llama 3.1 8B | Cheap and fast at high volume |
| Daily chat and coding | Llama 4 Maverick | Flagship multimodal, strong all-round |
| Very long documents | Llama 4 Scout | 10M context, fits one GPU |
| Fine-tuning on your data | Llama 3.1 8B or 70B | Small, cheap to train, easy to deploy |
| Private data, high volume | Maverick (self-host) | Full control, $0 per token |
| Images and documents | Llama 4 Scout or Maverick | Native multimodal input |
A simple rule helps. Prototype on a hosted provider, measure your real volume, then decide if self-hosting pays off. We worked with a team that ran everything on a frontier closed API. Moving routine work to a self-hosted Llama model and keeping only the hardest calls on a premium API cut their model bill by more than half with no drop in quality.
Older Llama models
Llama has shipped many versions, and older ones still run. If your code uses an older model, it will keep working, but the newer versions are stronger and often cheaper to serve. The table shows the main older models and what to use instead.
| Older model | Status | Move to |
|---|---|---|
| Llama 3.1 405B | Active, dense | Llama 4 Maverick |
| Llama 3.1 / 3.2 70B | Active, superseded | Llama 3.3 70B or Maverick |
| Llama 3, Llama 2 | Legacy | Llama 4 or Llama 3.3 |
| Code Llama | Retired for new work | Llama 4 Maverick |
A move is usually small. For a hosted provider you change the model name. For self-hosting you download the new checkpoint and update your serving config. The Meta Llama pages and the model cards on Hugging Face show every current and older version in one place, so check them before you migrate.
What this means for hiring
Knowing the lineup is step one. Getting value from it is the harder step. With Llama the biggest lever is the deployment choice. The teams that win know when to self-host, how to serve an MoE model efficiently, and how to route work between open weights and a hosted provider. That is engineering, not just a model choice.
Engineers who can deploy open models are scarce and costly in the United States. A senior AI engineer there runs $12,000 to $18,000 per month all in. The same skill from a pre-vetted engineer in Vietnam or the Philippines often costs $3,000 to $6,000 per month. The work is the same. The bill is not. Many of our clients build their whole open-model stack with an AI model deployment engineer from our network.
Compare other model families
Build your Llama stack with the right team
Llama gives you open weights, native multimodal input, and a record context window. The model is the easy part. The team that deploys and ships with it is what creates value. We match you with pre-vetted senior engineers across Asia in about 24 hours, with no upfront cost and payroll handled.
