Last updated:
Llama
Meta's open-weight model family — widely used as the base for countless fine-tuned and self-hosted AI products.
Model Facts
Latest model: Muse Glimmer 30B
- Provider
- Meta
- Released
- Apr 5, 2025
- Context window
- 1M tokens (Maverick) / 10M tokens (Scout)
- Weights
- Open
Benchmarks
80.5%
MMLU Pro
Maverick
69.8%
GPQA Diamond
Maverick
Full Model List
| Model | Released | Status | Note |
|---|---|---|---|
| Muse Glimmer 30B | Aug 10, 2026 | – | Meta's current open-weight model (Apache 2.0), distilled from Muse Spark |
| Muse Spark | Apr 8, 2026 | – | Meta's closed-weight frontier line — see its own entry |
| Llama 3.3 | Dec 6, 2024 | – | predecessor to Llama 4 |
| Llama 3.1 | Jul 23, 2024 | – | added a 405B parameter flagship size |
| Llama 3 | Apr 18, 2024 | – | major quality jump over Llama 2 |
| Code Llama | Aug 24, 2023 | legacy | code-specialized Llama 2 variants |
| Llama 2 | Jul 18, 2023 | – | first commercially-licensable Llama generation |
| Llama 4 Behemoth | – | – | previewed as still-in-training, never publicly released |
- Official model page:
- llama.com
- Official documentation:
- llama.com/docs/overview
- Other sources:
- Meta Muse Spark announcement
What it is
Llama is Meta's open-weight model family — widely used as the base for fine-tuned and self-hosted AI products. The latest Llama generation is Llama 4, released April 5, 2025 in two published sizes: Scout (10M-token context) and Maverick (1M-token context), both mixture-of-experts models. The larger Behemoth was previewed but never publicly released.
Where Meta is now
Meta's frontier work has moved to the closed-weight Muse line from Meta Superintelligence Labs (see Muse Spark). Meta has not formally discontinued Llama and existing models remain downloadable, but its newest open-weight release is Muse Glimmer 30B (August 10, 2026, Apache 2.0), distilled from Muse Spark.
Where these numbers come from
The benchmark figures are Meta's published Llama 4 Maverick results (MMLU Pro 80.5%, GPQA Diamond 69.8%) from the original release. They are dated compared with 2026 frontier models and are shown for the Llama family's own reference.
Notable traits
- Open weights under Meta's community license — free to download and run, with usage terms.
- Huge context windows for an open model (up to 10M tokens on Scout).
- No longer Meta's frontier line; see Muse Spark and Muse Glimmer.