What Do You Need to Run Meta Locally? VRAM by Version
Meta currently has 5 versions you can deploy locally. The smallest, Llama 3.2 3B, needs 5GB of VRAM at 4-bit; the largest, Muse Spark Open, is an MoE model and needs 14GB of VRAM plus about 495GB of RAM with MoE offloading. The cheapest device that gets one running is the Tesla V100 16GB, at about $370.
| Version | Released | Context | License | ||
|---|---|---|---|---|---|
| Muse Spark OpenMoE | Oct 15, 2026 | 816B | 42B | 1M | Muse Model License |
| Llama 4 Scout 17B 16EMoE | Apr 5, 2025 | 108.6B | 17B | 10M | Llama 4 Community License |
| Llama 3.3 70BDense | Dec 6, 2024 | 70.6B | 70.6B | 128K | Llama 3.3 Community License |
| Llama 3.2 3BDense | Sep 25, 2024 | 3.2B | 3.2B | 128K | Llama 3.2 Community License |
| Llama 3.1 8BDense | Jul 23, 2024 | 8B | 8B | 128K | Llama 3.1 Community License |
About Meta
Meta's open-weight models come in two lines: Llama, starting in 2023, and Muse, starting in 2026. Meta is headquartered in Menlo Park, California; Llama came out of FAIR, the AI research team Meta set up in 2013, and the Muse Spark Open model card is credited to Meta Superintelligence Labs.
The first-generation LLaMA, released in February 2023, was available only to researchers under a non-commercial license; commercial use has been allowed since Llama 2 in July of that year. The major versions since then are Llama 3 (April 2024); Llama 3.1 (July 2024), which added a 405B model and extended the context to 128K; Llama 3.2 (September 2024), which added 1B and 3B small models and 11B and 90B vision models; Llama 3.3 70B (December 2024); and Llama 4 (April 2025), the first in the series to use MoE and the first to be natively multimodal. Meta said cumulative Llama downloads passed 1 billion in March 2025. Llama uses Meta's own community license, which allows commercial use, but products with more than 700 million monthly active users must request a separate license, and the Open Source Initiative (OSI) considers that it does not meet the definition of open source.
As of October 2026 there has been no new Llama version since Llama 4, and Meta has moved on to Muse: the closed Muse Spark launched in April, the open-weight Muse Glimmer followed in August under Apache 2.0, and on October 15 Meta released Muse Spark Open, the open-weight version of Muse Spark, an 816B MoE that reads images, video and audio. It uses Meta's new Muse Model License, which allows commercial use, but a company that offers model API access and has more than 20 million US dollars in annual revenue must sign a separate agreement with Meta first.
- VendorMeta AI
- Versions5
- Parameters3.2B – 816B
- Latest versionMuse Spark Open
- License
- Llama 3.2 Community License / Llama 3.1 Community License / Llama 3.3 Community License / Llama 4 Community License / Muse Model License
- Official site
- ai.meta.com
- Hugging Face org
- meta-llama
How much VRAM do Meta models need at 2-bit?
2-bit, also called Q2, covers common GGUF files such as Q2_K and UD-Q2_K_XL. They are the smallest and lose the most quality, which makes them a way to get a model running first when VRAM is tight. Each row in the table below is a Meta version and each column is a context length; each cell is the total VRAM needed to deploy it at Q2 (weights + KV cache + 1.1GB runtime overhead).
| Version | WeightsGB | 32K contextGB | 128K contextGB | 256K contextGB | 1M contextGB |
|---|---|---|---|---|---|
| Muse Spark Open816B (42B active) | 311.2GGUF | 314.8 | 321.2 | 329.8 | 381.4 |
| Llama 4 Scout 17B 16E108.6B (17B active) | 42.4GGUF | 46.3 | 51.1 | 57.6 | 96.2 |
| Llama 3.3 70B70.6B | 27.4GGUF | 39.2 | 71.4 | — | — |
| Llama 3.2 3B3.2B | 1.3est. | 6.1 | 17.4 | — | — |
| Llama 3.1 8B8B | 3.7GGUF | 9.1 | 22 | — | — |
How much VRAM do Meta models need at 4-bit?
4-bit, also called Q4, corresponds to the widely used Q4_K_M files, which balance size and quality and are the default choice for local deployment. Each row in the table below is a Meta version and each column is a context length; each cell is the total VRAM needed to deploy it at Q4.
| Version | WeightsGB | 32K contextGB | 128K contextGB | 256K contextGB | 1M contextGB |
|---|---|---|---|---|---|
| Muse Spark Open816B (42B active) | 505.5GGUF | 509.1 | 515.5 | 524.1 | 575.7 |
| Llama 4 Scout 17B 16E108.6B (17B active) | 65.4GGUF | 69.3 | 74.1 | 80.6 | 119.2 |
| Llama 3.3 70B70.6B | 42.5GGUF | 54.3 | 86.5 | — | — |
| Llama 3.2 3B3.2B | 2GGUF | 6.8 | 18.1 | — | — |
| Llama 3.1 8B8B | 4.9GGUF | 10.3 | 23.2 | — | — |
How much VRAM do Meta models need at 8-bit?
8-bit, also called Q8, corresponds to Q8_0 files, which come closest to the original in quality and are also the largest. Each row in the table below is a Meta version and each column is a context length; each cell is the total VRAM needed to deploy it at Q8.
| Version | WeightsGB | 32K contextGB | 128K contextGB | 256K contextGB | 1M contextGB |
|---|---|---|---|---|---|
| Muse Spark Open816B (42B active) | 871.7GGUF | 875.3 | 881.7 | 890.3 | 941.9 |
| Llama 4 Scout 17B 16E108.6B (17B active) | 114.5GGUF | 118.4 | 123.2 | 129.7 | 168.3 |
| Llama 3.3 70B70.6B | 75GGUF | 86.8 | 119 | — | — |
| Llama 3.2 3B3.2B | 3.4GGUF | 8.2 | 19.5 | — | — |
| Llama 3.1 8B8B | 8.5GGUF | 13.9 | 26.8 | — | — |
FAQ
Are Llama and Muse open source and free for commercial use?
Every model listed here can be used commercially, but none of the licenses is MIT, Apache or anything similar; Meta wrote its own. Llama uses a community license, and each of the 4 Llama versions comes with its own copy, and the main terms are the same, with three you have to follow. First, a product with more than 700 million monthly active users in the calendar month before the version's release date must request a separate license from Meta. Second, if you distribute the model or a product built with it, you must include the license and prominently display “Built with Llama.” Third, if you train or fine-tune a model using Llama or its outputs and then release it, the model's name must begin with “Llama.” You also have to follow each version's Acceptable Use Policy, and that holds even if you download the model from Ollama or another source. The official repositories on Hugging Face ask you to fill in your contact information and accept the license first, and you can download only after your request is approved. Muse Spark Open uses a different license, the Muse Model License: it has no user-count or naming rules, and instead requires products to show “Built with Muse Spark Open” and companies offering model API access with more than 20 million US dollars in annual revenue to sign a separate agreement with Meta first.
Can you run Llama with Ollama?
Yes. All four Llama versions have official tags: llama3.2:3b, llama3.1:8b, llama3.3:70b and llama4:scout (which you can also write as llama4:16x17b). To pick a specific precision, use the full tag, such as llama3.3:70b-instruct-q8_0 or llama4:17b-scout-16e-instruct-q8_0. On Ollama, Llama 4 Scout can read images, while the three Llama 3.x models handle text only. Muse Spark Open can't run on Ollama yet: there's no official tag, and llama.cpp support for it hasn't been merged into mainline.
Can you use Llama 4 in the EU?
It depends on who you are. Llama 4's Acceptable Use Policy states that for the multimodal models in the release, individuals living in the EU and companies whose principal place of business is in the EU are not granted the rights to use, reproduce, distribute or modify them. Llama 4 Scout is a natively multimodal model, so it falls under this clause. The clause doesn't restrict end users: people in the EU can still use products or services that others have built with these models. Llama 3.2 3B, 3.1 8B and 3.3 70B, the other versions listed on this site, are text-only models and aren't affected. Muse Spark Open is multimodal too, but its license has no such regional restriction; individuals and companies in the EU get the same rights, licensed through Meta Platforms Ireland.
