Understand Llama 4 Maverick before you download it.
Llama 4 Maverick is Meta’s open-weights flagship: a mixture of experts with 17 billion parameters active per token, 128 experts in total and a one-million-token context window. It is the model teams download when the licence, not just the benchmark score, is the deciding factor — and the licence is the part most summaries leave out.
Every figure on this page comes from Meta’s published Llama 4 model card, the Llama 4 Community Licence and the Llama Models repository. There is no gallery: Meta publishes no static official figure for the Llama 4 family, and a page with no images is better than one padded with images we have no right to reproduce. Meta’s product navigation now leads with its Muse models, and the Llama 4 model page it used to serve is no longer reachable; the model card, the weights and the licence terms are unchanged.
How Meta describes Llama 4
Meta released Llama 4 in two sizes and describes both as natively multimodal mixture-of-experts models. Which one you pick depends on whether you are buying context or capability.
Two models, one collection
Scout is the context play: 17B activated across 16 experts, 109B parameters in total, a 10M-token window and a ~40T-token training corpus. Maverick is the capability play: 17B activated across 128 experts, 400B parameters in total, a 1M window and ~22T training tokens.
Multimodal by early fusion
Both fuse image and text inputs early rather than bolting a vision adapter onto a text model. The card lists multilingual text and images in, multilingual text and code out, and 12 supported languages: Arabic, English, French, German, Hindi, Indonesian, Italian, Portuguese, Spanish, Tagalog, Thai and Vietnamese.
22T tokens and a cutoff you should check
The card puts Maverick’s corpus at roughly 22T tokens with a knowledge cutoff of August 2024, and describes the models as static, trained on an offline dataset. Meta says future tuned models may follow as community feedback accumulates, but nothing after that date is in the weights you download today.
The licence is not an open-source licence
Llama 4 ships under the Llama 4 Community Licence Agreement, a custom commercial licence. It grants worldwide, non-exclusive, royalty-free use and modification, and then adds conditions: past 700 million monthly active users you must request a licence from Meta; you must display “Built with Llama” prominently; derivative models must carry a “Llama 4” name prefix. The Acceptable Use Policy applies on top.
Where the training data came from
Meta describes the corpus as a mix of publicly available and licensed data plus information from its own products and services — explicitly including publicly shared Instagram and Facebook posts and people’s interactions with Meta AI, with its Privacy Center linked for the detail. That disclosure is in the model card, not an inference from it.
How the Llama models compare
Both Llama 4 rows come from the model card Meta publishes; the two older models are included to show the size of the jump.
| Model | Architecture | Context length |
|---|---|---|
| Llama 4 Maverick | MoE — 17B activated of 400B, 128 experts | 1M |
| Llama 4 Scout | MoE — 17B activated of 109B, 16 experts | 10M |
| Llama 3.1 405B | Dense — 405B | 128K |
| Llama 3.3 70B | Dense — 70B | 128K |
Meta’s own benchmarks put Maverick ahead of Llama 3.1 405B on MMLU (85.5 against 85.2), MMLU-Pro (62.9 against 61.6), MATH (61.2 against 53.5) and MBPP (77.6 against 74.4) — at a similar total parameter count, but with only 17B parameters active per token.
What the model card documents
Meta publishes specifications and benchmark tables rather than example images, so this section carries the capability claims the card actually makes.
Native image reasoning
Maverick scores 73.4 on MMMU, 59.6 on MMMU Pro and 73.7 on MathVista in Meta’s instruction-tuned table, and 85.3 on ChartQA and 91.6 on DocVQA in the pre-trained table.
A million-token window
1M for Maverick, 10M for Scout, both stated in the model card. Meta publishes no retrieval benchmark to go with those numbers, so treat them as capacity rather than as a measured long-context capability.
Twelve supported languages
Arabic, English, French, German, Hindi, Indonesian, Italian, Portuguese, Spanish, Tagalog, Thai and Vietnamese. Pre-training spanned 200 languages, and the licence allows fine-tuning into languages beyond the twelve.
Mixture-of-experts efficiency
17B parameters are activated per token out of 400B in total. That is the point of the architecture: near-large-model quality at a fraction of the per-token compute, provided you have the memory to hold the experts.
Fine-tuning and distillation are allowed
The card names synthetic data generation, distillation and adapting the pretrained models among the supported use cases, and the licence grants derivative works — subject to the naming, attribution and acceptable-use conditions.
Training cost, disclosed
Meta publishes the training footprint: 2.38M H100-80GB GPU hours for Maverick and 5.0M for Scout, 7.38M in total, with 1,999 tons of location-based CO2 equivalent offset to zero on a market basis. Few vendors put this in a model card.
Documented specifications
Every row below is stated on Meta’s published Llama 4 model card or in the Llama 4 Community Licence Agreement.
- Developer
- Meta
- Model
- Llama 4 Maverick (17B×128E)
- Also in the collection
- Llama 4 Scout (17B×16E)
- Total parameters
- 400B (Maverick) · 109B (Scout)
- Activated parameters
- 17B per token
- Experts
- 128 (Maverick) · 16 (Scout)
- Context length
- 1M (Maverick) · 10M (Scout)
- Input modalities
- Multilingual text and image
- Output modalities
- Multilingual text and code
- Training corpus
- ~22T tokens (Maverick) · ~40T (Scout)
- Knowledge cutoff
- August 2024
- Release date
- 5 April 2025
- Licence
- Llama 4 Community Licence
Documented access channels
Everything Llama 4 gives you arrives as a download: the weights, the model card, the licence and the CLI that fetches them. There is no hosted Meta endpoint for this model.
Llama 4 model card
The specification table, both benchmark tables, the intended-use statement, the training footprint and the safety notes — the source for every figure on this page.
OpenLlama 4 Community Licence
The binding terms: the 700-million-user clause, the “Built with Llama” display requirement and the attribution notice you must ship in a copy of the materials.
OpenLlama Models repository
Meta’s own repository for the whole family, with the model table across versions and the CLI documentation.
OpenWeights on Hugging Face
Published under Meta’s organisation on Hugging Face. The repositories are gated: you request access, accept the licence, and receive a signed download URL that expires after 24 hours.
OpenLlama 4 Maverick questions
Is Llama 4 open source?
No, and the distinction is worth keeping. The weights are downloadable and the licence is royalty-free for most users, but it is the Llama 4 Community Licence Agreement — a custom commercial licence rather than an OSI-approved one. It adds a 700-million-user threshold, a “Built with Llama” display rule and a naming rule for derivative models.
What exactly does the 700-million-user clause say?
If the monthly active users of the products or services made available by you or your affiliates exceeded 700 million in the previous calendar month as of the model’s release date, you must request a licence from Meta and may not exercise any rights under the agreement until Meta grants one.
How much context does it have?
1M tokens for Maverick and 10M for Scout, both from Meta’s model card. Meta publishes no retrieval benchmark alongside those figures, so they describe capacity rather than measured long-context accuracy.
When was it released, and how current is it?
Llama 4 was released on 5 April 2025 with a knowledge cutoff of August 2024, and Meta describes the models as static, trained on an offline dataset. Meta’s newer frontier work sits under the Muse brand and its product navigation lists those models first; Llama 4 remains published, downloadable and licensed under the same terms.
Can I fine-tune it?
Yes. The model card names synthetic data generation, distillation and adapting the pretrained checkpoints among the supported use cases, and the licence grants derivative works — with the naming, attribution and acceptable-use conditions attached.
Can I run it locally?
Maverick holds 400B parameters, and the memory cost follows the total rather than the 17B active per token. Meta publishes no minimum hardware figure; the card documents the training configuration, not a deployment one. This is not a laptop model.
Does it support images?
Yes, natively and from early in pre-training rather than through a bolt-on adapter. For Maverick the card reports MMMU at 73.4 and MMMU Pro at 59.6, against 69.4 and 52.2 for Scout.