Meta Llama: Versions, License, How to Run & Polish Test

👁 128 reads

Llama is Meta’s family of open-weight language models: you can download and run them locally, including for commercial use, under the Llama Community License. The latest Llama (Llama 4 Scout and Maverick) launched in April 2025, and since 2026 Meta has been releasing new models under the Muse brand. On home computers, people most commonly run Llama 3.1 8B and 3.2 3B today - they handle simple tasks in Polish correctly, but Polish isn’t among the officially supported languages, as our test shows.

What is Llama

Llama (originally LLaMA, standing for “Large Language Model Meta AI”) is a family of language models from Meta, the company behind Facebook, Instagram, and WhatsApp. The first version from February 2023 was available only to researchers, and its weights leaked online within a week. Starting with Llama 2 (July 2023), Meta has made the weights available to anyone who accepts the license. We explain why the company gives away models for free in our article Mark Zuckerberg and AI.

For three years, Llama was the reference point for all open models. Thousands of specialized versions were fine-tuned from it, and tools for running AI locally, such as llama.cpp and Ollama, were built around it - hence “llama” in their names. In 2026, the situation changed: Meta moved development to a new lab called Meta Superintelligence Labs, in April 2026 replaced Llama in its chat products with the closed Muse Spark model, and released its new open model under the name Muse Glimmer.

All Llama versions

Muse: what Meta is doing after Llama

Muse Glimmer is the practical successor to Llama for people who want an open Meta model. It has 30 billion parameters, 128K token context, understands images and video, and takes 18 GB in Ollama, so it fits on a 24 GB card. Unlike Llama 4, it has an Apache 2.0 license without European Union exceptions. We covered the Muse Spark model and its Polish language results in our article Muse Spark 1.3.

How to get started with Llama step by step

Easiest: locally in Ollama

1. Download and install Ollama from ollama.com.
2. Check how much memory your graphics card has (in Windows: Task Manager, Performance tab, GPU).
3. Up to 4 GB: type ollama run llama3.2:3b (2 GB). 8 GB or more: ollama run llama3.1:8b (4.9 GB).
4. Chat in the terminal or install a GUI wrapper like LM Studio, which downloads the same models from Hugging Face.
5. Write prompts in Polish, but for important texts add “Respond in correct Polish” - Llama sometimes switches to English or uses English syntax patterns.

A complete guide with hardware requirements, our speed measurements, and version selection is in the article Llama locally.

Without installation

• Meta AI in WhatsApp, Messenger, Instagram, and at meta.ai has been available in Poland since March 2025, but Meta doesn’t specify which model - as of April 2026 in the US it’s Muse Spark, not Llama.
• Through API from providers like OpenRouter, Together, or Groq. This is the cheapest way to chat with large versions (Llama 3.3 70B, Llama 4 Maverick) that you can’t run at home.

Llama license: what you can and can’t do

Llama is not open source by the Open Source Initiative definition. It’s “open weights” under Meta’s own license. In practice:

• You can: download the model, run it locally, fine-tune it, and embed it in a paid product without fees.
• You must: accept the license and Acceptable Use Policy (AUP), and include a “Built with Llama” notice in products based on Llama.
• You can’t without separate Meta approval: use Llama in a product that had over 700 million monthly active users on the model’s release date.
• EU trap: Llama 4’s use policy states that rights to multimodal models “are not granted” to individuals residing in the EU and companies headquartered in the EU. The same restriction applies to Llama 3.2 vision models (11B and 90B). Exception: regular users of a finished product built on these models. Text-only models (Llama 3.1, 3.2 1B and 3B, 3.3 70B) don’t have this restriction.

Conclusion for EU companies: if you want an open Meta model for your own product, choose Llama 3.1 or 3.3, or the new Muse Glimmer on Apache 2.0, not Llama 4. This is our interpretation of the license, not legal advice.

How much does Llama cost

The weights themselves are free. You pay for hardware and electricity or for API from an external provider. Lowest rates on OpenRouter, October 3, 2026, per million tokens:

Our entire five-prompt test cost $0.0006 on Llama 3.3 70B and $0.0006 on Llama 4 Maverick. Muse Glimmer used $0.0101 because it “thought” extensively before each response (2.6 to 12.1 thousand characters of reasoning per task; for the summary I had to raise the token limit because the first time it used the entire limit on thinking and didn’t return an answer).

Llama in Polish

Meta officially supports eight languages in Llama 3.1 (English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai), and twelve in Llama 4. Polish isn’t on either list. The model understands Polish because it saw it in training data, but text quality depends heavily on size:

• Llama 3.2 1B and 3B - we didn’t test these; since Llama 3.1 8B makes quite a few errors, we wouldn’t expect better Polish from smaller models. This is our opinion.
• Llama 3.1 8B - understandable Polish with visible calques and errors, details in our test below.
• Llama 3.3 70B and Llama 4 Maverick - correct Polish, but without flair; in our test, 70B summarized text too briefly, and in the email confused the form of address.

If you care about Polish text from a local model, Bielik and Gemma 4 12B perform better. A comparison of seven local models on 13 tasks is in the article local AI models in Polish.

Our test: 5 prompts in Polish

On October 3, 2026, we gave five prompts in Polish (summary with word limit, email to administrator, correction of 10 errors, explanation for a child, translation of idioms) to four Meta models: Llama 3.1 8B locally in Ollama on RTX 4090, and Llama 3.3 70B, Llama 4 Maverick, and Muse Glimmer 30B through OpenRouter. The same pool included Bielik, Mistral models, and Gemma 4. Responses were blindly evaluated by Claude Opus 5.5 on a 1-10 scale.

For comparison: local Gemma 4 12B scored 8.33, Bielik 11B got 7.60, and the best Mistral (Large 3) scored 7.67. Muse Glimmer won the entire test, so the open successor to Llama is currently Meta’s best model for Polish texts among those we tested.

What we remember from the responses:

• Llama 3.1 8B started a translation with an awkward phrase, left an error in the correction task, and explained to a child that “the sky reflects light.” In the email it used an incorrect form of address and added a fact that wasn’t in the prompt.
• Llama 3.3 70B summarized an article in 17 words, losing the city, date, and vote result.
• Llama 4 Maverick did a flawless correction, but translated “missed the deadline by a hair” as the opposite meaning, and mentioned “Rayleigh scattering” to an eight-year-old.
• Muse Glimmer was the only Meta model to correctly translate the idiom, but thought extensively before each response, costing 16 times more than Maverick.

Locally, Llama 3.1 8B generated 149 tokens per second and used 5.9 GB of card memory - it’s fast, but too weak for Polish texts. Full results: Bielik, Llama or Mistral in Polish.

Llama compared to other open models

We covered the European alternative, Mistral models, in our guide Mistral AI, and the Chinese alternative in the article about DeepSeek.

Who is Llama for

• Learning local AI - Llama 3.1 8B and 3.2 3B are the best-documented models online, with ready-made tutorials for every tool.
• English tasks - summaries, classification, extracting data from text.
• Fine-tuning - Llama has the most ready-made recipes and derivative versions.
• Not for you if you write mainly in Polish (choose Bielik or Gemma) or are building a vision product in the EU on Llama 4 (choose Muse Glimmer or Mistral).

Frequently asked questions

Is Llama free?

Yes. You download Llama weights for free and can use them commercially, as long as your product doesn’t exceed 700 million monthly active users and you follow Meta’s use policy. You only pay for hardware or API from an external provider.

What is the latest version of Llama?

Llama 4 Scout and Maverick from April 5, 2025. Since then, Meta hasn’t released a new Llama - new models are Muse Spark (closed) and Muse Glimmer 30B (open, Apache 2.0, August 2026).

Does Llama work in Polish?

It understands and writes in Polish, but Polish isn’t an officially supported language. Small Llama 3.1 8B makes noticeable errors, large versions (70B, Maverick) write correctly, though stiffly.

Can I use Llama 4 in the EU?

As a user of a finished application - yes. As an individual or company from the EU that wants to download and deploy Llama 4 themselves, Meta’s license doesn’t grant you rights to multimodal models, and Llama 4 Scout and Maverick are exactly that. It’s safer to choose Llama 3.1, 3.3, or Muse Glimmer.

What computer do you need for Llama?

Llama 3.2 3B works even on a laptop without a graphics card. Llama 3.1 8B needs a card with 6-8 GB memory, Llama 3.3 70B - about 43 GB memory (two 24 GB cards or a Mac with large memory). Details in the article Llama locally.

Llama or ChatGPT?

They’re different things. ChatGPT is a ready-made cloud service with much more powerful models. Llama is a model you run yourself: weaker, but free, private, and works offline.

What is Muse Glimmer?

It’s an open Meta model from August 10, 2026: 30 billion parameters, Apache 2.0 license, image and video support, 18 GB in Ollama. In practice, the successor to Llama for people who want to run a Meta model locally.

Is Meta AI in WhatsApp using Llama?

Until April 2026, Meta AI ran on Llama. Since April 2026 in the US, it’s powered by Muse Spark. For Poland, Meta doesn’t specify which model responds.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Sources and verification date

• Wikipedia: Llama (language model) (dates and version sizes)
• Llama 4 announcement and Llama 3.1 on Meta AI blog
• Llama 4 Acceptable Use Policy (EU clause)
• Muse Glimmer in Ollama library and llama3.1, llama3.2, and llama4 pages (file sizes)
• Llama models on OpenRouter (API pricing)
• Our own 5-prompt test: Ollama 0.35.1 on RTX 4090 and OpenRouter, October 3, 2026

Verified October 3, 2026.

Kinetyra