Technical profile
What developers should know
Architecture and modalities
Maverick is an autoregressive mixture-of-experts model. Meta uses early fusion to combine text and image inputs within the model instead of treating vision as a separate downstream feature.[1]
The published model card lists multilingual text and images as inputs and multilingual text and code as outputs. The 12 supported languages are Arabic, English, French, German, Hindi, Indonesian, Italian, Portuguese, Spanish, Tagalog, Thai, and Vietnamese.[1]
Training and data scope
Meta reports approximately 22 trillion pretraining tokens for Maverick. Its disclosed data mix includes publicly available and licensed data plus information from Meta products and services; the pretraining-data cutoff is August 2024.[1]
The one-million-token context window is a model specification. Hosted providers can expose different practical limits, so production integrations should verify the endpoint-specific input, output, image, and rate limits separately.[1]
Intended uses and limits
Meta positions the instruction-tuned checkpoint for assistant-style chat and visual reasoning. The model card also names visual recognition, image reasoning, captioning, and general image question answering as intended vision uses.[1]
Meta reports testing image understanding with as many as five input images. Use beyond the listed languages or capabilities requires deployment-specific evaluation and safeguards.[1]
Choose the artifact
Official checkpoints
This page covers the Maverick family. Select the checkpoint that matches your serving stack and whether you need a pretrained or instruction-tuned model.[1] [2]
meta-llama/Llama-4-Maverick-17B-128E
Pretrained checkpoint for adaptation and natural-language generation.[1]
↗ Instruct · Transformersmeta-llama/Llama-4-Maverick-17B-128E-Instruct
Instruction-tuned checkpoint for chat and visual-reasoning applications.[2]
↗ Instruct · original formatmeta-llama/Llama-4-Maverick-17B-128E-Instruct-Original
Original checkpoint format for Meta’s llama-stack codebase.[3]
↗Meta releases Maverick in BF16 and FP8-quantized forms and reports that the FP8 weights fit on one H100 DGX host. Treat that as a vendor deployment reference, not a memory guarantee for every serving stack, batch size, or context length.[2]
Reported evaluation
Selected instruction-tuned results
Vendor-reported evaluations from Meta’s model card. All listed evaluations used BF16 models; scores are not independent llms.help tests.[2]
| Category | Benchmark | Evaluation | Score |
|---|---|---|---|
| Image reasoning | MMMU | 0-shot accuracy | 73.4 |
| Coding | LiveCodeBench | 0-shot pass@1 | 43.4 |
| Reasoning & knowledge | MMLU Pro | 0-shot macro average accuracy | 80.5 |
| Reasoning & knowledge | GPQA Diamond | 0-shot accuracy | 69.8 |
| Multilingual | MGSM | 0-shot average exact match | 92.3 |
Before deployment
Llama 4 Community License
Maverick uses Meta’s custom Llama 4 Community License, which includes its own redistribution, attribution, acceptable-use, and additional commercial terms. Review those terms before shipping.[4]
Read the license ↗