Llama review: Meta's open-weight models & how to run them

Llama is Meta's family of large language models, whose weights were released openly, which is why it played a big part in the rise of open-source AI. Thousands of specialized models, from coding assistants to models fine-tuned for different languages, are built on Llama, and almost every local-running tool supports it. Llama is not a chatbot; it is a model that you run yourself with tools such as Ollama or LM Studio, or call through API providers. Meta has moved its focus to new, mostly closed models and the family no longer gets new generations, but the released models remain available and widely used.

What is Llama good for?#

Internal company chatbot: a company that does not want its data leaving its servers runs Llama on an internal server and connects it to its own documents.
Fine-tuning base: a research team gives a small Llama model extra training on specialist data, such as legal text.
Classification and extraction: for simple jobs like sorting tickets or pulling names and dates out of English text, a small cheap model is enough.
Learning and research: a student studies the structure and behavior of a real language model.
Offline use: on a laptop with no internet, for summarizing and drafting English text.

Strengths and weaknesses#

Strengths:

A very large ecosystem, with plenty of tools, tutorials and derived models.
Many sizes, from light models for laptops to large ones for servers.
Supported in every local-running tool and most API providers.

Weaknesses:

Weaker than open competitors such as Gemma and Qwen in several non-English languages.
No longer gets new generations, and open rivals have pulled ahead.
The license is not open source in the strict sense and has conditions you should read.

Free tier and pricing#

The model weights are free, and running locally costs nothing beyond hardware. If you have no hardware, various API providers offer Llama at usage-based prices, usually cheap. Check the provider's own site for current pricing.

Language support#

Llama understands text in many languages, but when writing in languages other than English, especially in small sizes, it can produce incorrect sentences and made-up words, so it is not our first choice for non-English work. Once the model files are downloaded, it runs fully offline. Keep in mind that the Llama license requires compliance with export regulations.

Llama alternatives#

Gemma: Google's open models, with a more permissive license and better small-size multilingual quality.
Qwen: Alibaba's open family, strong in programming and multilingual use.
DeepSeek: strong open models for reasoning and code.
Le Chat (Mistral): European models, some released under open licenses.

To get started, read what a large language model (LLM) is and how it learns, open-source AI models: what they are and when they help and running AI on a personal computer.

Frequently asked questions about Llama

Is Llama free?

Yes. The model weights are released for free and you can run them on your own machine. Use is governed by Meta's own license, which does not restrict most users and businesses.

Is Llama open source?

Not in the strict sense. The weights are openly available, but the license adds conditions such as an acceptable-use policy and limits for very large companies, which is why it is called an "open-weight" model.

Is Llama still being updated?

Meta has shifted its focus to new, mostly closed models, and the Llama family is not getting a new generation. The released models are still available to download and use.

Does Llama work well in languages other than English?

It understands many languages, but its writing in them, especially in the smaller sizes, can be weak and error-prone. For non-English work, models such as Gemma or Qwen are often better.

Alternatives to Llama