benchmarks
Meta's Llama 4 Scout and Maverick are downloadable-weight, mixture-of-experts models with native image-and-text understanding and text output, distributed under the Llama 4 Community License. Documented context capacities are 10 million tokens for Scout and one million for Maverick; the benchmark figures below are Meta's April 2025 Maverick results, not family-wide scores.
Benchmark results
Every figure links to where it was published. Compare like with like — harnesses and prompt setups differ between vendors.
| Benchmark | Score | Source |
|---|---|---|
| AIMECompetition mathematics | 61.2%Apr 2025 | View |
| GPQA DiamondPhD-level science questions | 69.8%Apr 2025 | View |
| LiveCodeBenchFresh competitive-programming problems | 43.4%Apr 2025 | View |
| MMLU-ProBroad expert knowledge and reasoning | 80.5%Apr 2025 | View |
| MMMUMultimodal (image + text) understanding | 73.4%Apr 2025 | View |
Weighted mean of 5 verified benchmark results: AIME 61.2%, GPQA Diamond 69.8%, LiveCodeBench 43.4%, MMLU-Pro 80.5%, MMMU 73.4%.
Capabilities
- Native image-and-text understanding with text responses, including visual question answering and document interpretation.
- Long-context text processing, with documented capacities of 10 million tokens for Scout and one million for Maverick.
- Text support for twelve explicitly documented languages.
- Code generation and mathematical and scientific question answering, supported by published vendor evaluations.
- Downloadable weights for deployment and customization under the Llama 4 Community License.
- Mixture-of-experts inference with approximately 17 billion active parameters per token in both variants.
What improved
- Expands explicitly supported text languages from eight in Llama 3.1 to twelve.
- Introduces mixture-of-experts architecture instead of the dense architecture used by Llama 3 models.
- Uses native multimodal early fusion rather than Llama 3.2 Vision's adapter-based approach.
- Extends documented context capacity beyond Llama 3.1's 128K-token window.
Known flaws & limitations
- Performance and safety outside the twelve explicitly supported languages require additional validation.
- Sparse activation does not eliminate storage and memory requirements for approximately 109B total Scout parameters or 400B Maverick parameters.
- The custom license restricts use and distribution, including multimodal-model grants to EU-based individuals and companies, with an end-user exception.
- Generates text rather than native audio, video or image outputs.
- The documented context window is a capacity specification, not a guarantee of reliable retrieval or reasoning across every token.
Risks, safety & security issues
- Outputs can be inaccurate or fabricated; consequential factual claims require independent verification.
- Safety tuning does not guarantee prevention of harmful outputs, jailbreaks or instruction-following failures.
- Self-hosted deployments require application-specific safeguards, security testing and abuse monitoring.
- Connecting the model to tools or confidential data requires independent access controls and validation of model-generated actions.
Good fit for
- Image-and-text assistants that produce textual explanations or answers.
- Long-document analysis and retrieval-assisted workflows with source checking.
- Multilingual assistants operating within the twelve documented text languages.
- Customizable, self-hosted coding and knowledge assistants where licensing and infrastructure requirements are acceptable.
Poor fit for
- Unsupervised medical, legal, financial or other consequential decision-making.
- Applications requiring native image, audio or video generation.
- Deployments requiring guaranteed factual accuracy or reliable reasoning across every token of the maximum context.
- Resource-constrained deployments that cannot accommodate the full model weights and context-related memory demands.
References
Profile last reviewed 10/21/2026 and re-checked automatically every few hours against the vendor's published documentation, papers and public leaderboards. Scores come from the cited benchmark results above — confirm limits and pricing against the provider's own documentation before building on them.