The Llama 4 model card is the release document Meta published on April 5, 2025 for Llama 4 Scout and Llama 4 Maverick, the first Llama models to use a mixture-of-experts architecture and native early-fusion multimodality. It is maintained in Meta's llama-models repository and covers model specifications, intended and out-of-scope use, training compute and emissions, training data, benchmark results, quantization, safeguards and critical-risk areas, and limitations. It is a product card rather than a technical report: it states results and describes process, without the methods detail of The Llama 3 Herd of Models.
Models and specifications
Both released models activate 17B parameters per token. Llama 4 Scout is described as 17B activated of 109B total with 16 experts, a 10M-token context length and roughly 40 trillion pre-training tokens; Llama 4 Maverick as 17B activated of 400B total with 128 experts, a 1M-token context length and roughly 22 trillion tokens. Both take multilingual text and image input and produce multilingual text and code, with a knowledge cutoff of August 2024. The card lists twelve supported languages — Arabic, English, French, German, Hindi, Indonesian, Italian, Portuguese, Spanish, Tagalog, Thai and Vietnamese — while noting that pre-training covered 200 languages, and states that image understanding was tested for up to five input images.
Pre-training used 7.38M H100-80GB GPU hours (5.0M for Scout, 2.38M for Maverick), which the card reports as 1,999 tons CO2eq of location-based emissions and zero tons market-based. Training data is described as "a mix of publicly available, licensed data and information from Meta's products and services," including publicly shared Instagram and Facebook posts and users' interactions with Meta AI. Scout is released in BF16 with on-the-fly int4 quantization code; Maverick is released in both BF16 and FP8. The card states that all reported evaluations were run on BF16 models.
Benchmark results
Results are reported in two tables, both comparing the new models against Meta's own previous generation rather than against other developers' models. The pre-trained table compares Llama 3.1 70B, Llama 3.1 405B, Llama 4 Scout and Llama 4 Maverick on MMLU, MMLU-Pro, MATH, MBPP, TydiQA, ChartQA and DocVQA, with each row's shot count and metric stated; the Llama 3 columns are marked "No multimodal support" on the image rows. The instruction-tuned table compares Llama 3.3 70B, Llama 3.1 405B, Scout and Maverick on MMMU, MMMU Pro, MathVista, ChartQA, DocVQA (test), LiveCodeBench, MMLU Pro, GPQA Diamond, MGSM and MTOB, with the Llama 3 columns marked "Context window is 128K" on the MTOB long-context rows. The card notes that its MMMU Pro figure is the average of the Standard and Vision tasks, and that its LiveCodeBench window is 10/01/2024–02/01/2025.
No comparison to non-Meta models appears anywhere in the card, and no human-preference or arena results are reported.
Safeguards and critical risks
The safety section describes a three-part strategy — model-level fine-tuning, system-level protections, and community tooling — and states that Meta deliberately worked to reduce refusals of benign prompts and to remove "preachy and overly moralizing language" from refusal tone, including a sample system prompt instructing the model not to lecture users and not to refuse prompts about political and social issues. System-level protections named are Llama Guard, Prompt Guard and Code Shield, which the card says developers should deploy alongside the model.
Three critical-risk areas are described: CBRNE helpfulness, assessed with expert-designed evaluations of whether the model could meaningfully increase malicious actors' capabilities; child safety, addressed through pre-training data filtering, expert assessment and expanded benchmarks covering multi-image and multilingual capabilities; and cyber-attack enablement, where the card reports the conclusion that "Llama 4 models do not introduce risk plausibly enabling catastrophic cyber outcomes." None of the three sections reports a numeric result, an evaluation threshold, or a named dangerous-capability benchmark.
Licensing and use restrictions
The models are released under the Llama 4 Community License, a custom commercial licence. The card lists as out of scope any use violating applicable law or the Acceptable Use Policy, and any use "in languages or capabilities beyond those explicitly referenced as supported in this model card," while permitting fine-tuning into additional languages under the licence.
Provenance and confidence
Confidence is high for the specifications and the reported figures, which come directly from Meta's card. All benchmark numbers are developer-reported and are compared only against Meta's own earlier models, so the card supplies no cross-developer comparison. The card also carries none of the safety numbers that its critical-risk framing describes, which limits what can be checked against it. Independent reproductions of Meta's Llama 4 figures reported lower scores than the card, and the separately reported submission of an unreleased "experimental chat version" to LMSYS Chatbot Arena is documented on Llama 4 (Meta) rather than in this card.
Reported results
Evaluation results the document states for which there is no separate benchmark page; results on benchmarks that have a page are recorded there and cite this page. Scores and settings are as printed in the document.
| Evaluation | Model | Score | Setting | Section |
|---|---|---|---|---|
| MBPP — Pre-trained | Llama 3.1 70B | 66.4 | 3-shot, pass@1, bf16; developer-reported | Benchmarks — Pre-trained models |
| MBPP — Pre-trained | Llama 3.1 405B | 74.4 | 3-shot, pass@1, bf16; developer-reported | Benchmarks — Pre-trained models |
| MBPP — Pre-trained | Llama 4 Scout | 67.8 | 3-shot, pass@1, bf16; developer-reported | Benchmarks — Pre-trained models |
| MBPP — Pre-trained | Llama 4 Maverick | 77.6 | 3-shot, pass@1, bf16; developer-reported | Benchmarks — Pre-trained models |
| TydiQA — Pre-trained | Llama 3.1 70B | 29.9 | 1-shot, average/f1, bf16; developer-reported | Benchmarks — Pre-trained models |
| TydiQA — Pre-trained | Llama 3.1 405B | 34.3 | 1-shot, average/f1, bf16; developer-reported | Benchmarks — Pre-trained models |
| TydiQA — Pre-trained | Llama 4 Scout | 31.5 | 1-shot, average/f1, bf16; developer-reported | Benchmarks — Pre-trained models |
| TydiQA — Pre-trained | Llama 4 Maverick | 31.7 | 1-shot, average/f1, bf16; developer-reported | Benchmarks — Pre-trained models |
| ChartQA — Pre-trained | Llama 4 Scout | 83.4 | 0-shot, relaxed_accuracy, bf16; Llama 3 columns marked "No multimodal support"; developer-reported | Benchmarks — Pre-trained models |
| ChartQA — Pre-trained | Llama 4 Maverick | 85.3 | 0-shot, relaxed_accuracy, bf16; developer-reported | Benchmarks — Pre-trained models |
| DocVQA — Pre-trained | Llama 4 Scout | 89.4 | 0-shot, anls, bf16; Llama 3 columns marked "No multimodal support"; developer-reported | Benchmarks — Pre-trained models |
| DocVQA — Pre-trained | Llama 4 Maverick | 91.6 | 0-shot, anls, bf16; developer-reported | Benchmarks — Pre-trained models |
| MathVista — Instruction tuned | Llama 4 Scout | 70.7 | 0-shot, accuracy, bf16; developer-reported | Benchmarks — Instruction tuned models |
| MathVista — Instruction tuned | Llama 4 Maverick | 73.7 | 0-shot, accuracy, bf16; developer-reported | Benchmarks — Instruction tuned models |
| ChartQA — Instruction tuned | Llama 4 Scout | 88.8 | 0-shot, relaxed_accuracy, bf16; developer-reported | Benchmarks — Instruction tuned models |
| ChartQA — Instruction tuned | Llama 4 Maverick | 90.0 | 0-shot, relaxed_accuracy, bf16; developer-reported | Benchmarks — Instruction tuned models |
| DocVQA (test) — Instruction tuned | Llama 4 Scout | 94.4 | 0-shot, anls, bf16; developer-reported | Benchmarks — Instruction tuned models |
| DocVQA (test) — Instruction tuned | Llama 4 Maverick | 94.4 | 0-shot, anls, bf16; developer-reported | Benchmarks — Instruction tuned models |
| MGSM — Instruction tuned | Llama 3.3 70B | 91.1 | 0-shot, average/em, bf16; developer-reported | Benchmarks — Instruction tuned models |
| MGSM — Instruction tuned | Llama 3.1 405B | 91.6 | 0-shot, average/em, bf16; developer-reported | Benchmarks — Instruction tuned models |
| MGSM — Instruction tuned | Llama 4 Scout | 90.6 | 0-shot, average/em, bf16; developer-reported | Benchmarks — Instruction tuned models |
| MGSM — Instruction tuned | Llama 4 Maverick | 92.3 | 0-shot, average/em, bf16; developer-reported | Benchmarks — Instruction tuned models |
| MTOB (half book) eng→kgv — Instruction tuned, long context | Llama 4 Scout | 42.2 chrF | Llama 3 columns marked "Context window is 128K"; developer-reported | Benchmarks — Instruction tuned models |
| MTOB (half book) kgv→eng — Instruction tuned, long context | Llama 4 Scout | 36.6 chrF | developer-reported | Benchmarks — Instruction tuned models |
| MTOB (half book) eng→kgv — Instruction tuned, long context | Llama 4 Maverick | 54.0 chrF | developer-reported | Benchmarks — Instruction tuned models |
| MTOB (half book) kgv→eng — Instruction tuned, long context | Llama 4 Maverick | 46.4 chrF | developer-reported | Benchmarks — Instruction tuned models |
| MTOB (full book) eng→kgv — Instruction tuned, long context | Llama 4 Scout | 39.7 chrF | developer-reported | Benchmarks — Instruction tuned models |
| MTOB (full book) kgv→eng — Instruction tuned, long context | Llama 4 Scout | 36.3 chrF | developer-reported | Benchmarks — Instruction tuned models |
| MTOB (full book) eng→kgv — Instruction tuned, long context | Llama 4 Maverick | 50.8 chrF | developer-reported | Benchmarks — Instruction tuned models |
| MTOB (full book) kgv→eng — Instruction tuned, long context | Llama 4 Maverick | 46.7 chrF | developer-reported | Benchmarks — Instruction tuned models |
Relationships
- supports: Llama 4 (Meta) (canonical primary reference)
- supersedes: The Llama 3 Herd of Models (Meta, July 2024) (as the release document for the next Llama generation; it is far less detailed)
- related: Meta AI, Open-Source AI / Open-Weight Models, Llama 3 (Meta)
- instance-of: AI System Cards