Industry / Models / Meta
Meta / Llama
An open-weight multimodal MoE model for teams that need control over inference, adaptation, and deployment.
- Model snapshot
- Llama 4 Maverick 17B-128E Instruct
- Deployment
- Self-hosted or managed inference for the released checkpoint
- Context & outputs
- Maverick: 1M-token model-card context
- Access & license
- Llama 4 Community License; not Apache or MIT
Architecture and capabilities
Llama 4 Maverick uses a mixture-of-experts architecture with 400B total parameters and 17B activated parameters. It accepts text and images and generates text and code. The model card describes a one-million-token context. Meta model card.
Openness and deployment
Maverick is an open-weight release under the custom Llama 4 Community License. That differs from a permissive Apache or MIT license. Its published benchmarks belong to this April 2025 checkpoint and its stated evaluation setup.
Our engineering approach
We consider Llama when organizations want control over inference infrastructure and model adaptation. The deployment decision includes license review, memory sizing, quantization quality, and application-specific evaluations. For document and image workflows, we measure extraction correctness and multilingual behavior. For coding, we run repository-level tests as well as small code-generation tasks: a competitive-programming score alone does not establish whether an agent can safely maintain an application.
Published evidence
Benchmark snapshot.
Historical April 2025 Maverick checkpoint, not a claim about Meta’s latest model. LiveCodeBench uses the October 2024–February 2025 slice, zero-shot pass@1. Other rows use the model-card metrics.
| Parameter / benchmark | Result | Evaluation setting | Source |
|---|---|---|---|
| Code generationLiveCodeBench | 43.4% | Oct 2024–Feb 2025; zero-shot pass@1 | Published result |
| Scientific reasoningGPQA Diamond | 69.8% | Zero-shot accuracy | Published result |
| Knowledge and reasoningMMLU Pro | 80.5% | Zero-shot macro-average accuracy | Published result |
| Image understandingChartQA | 90.0% | Relaxed accuracy | Published result |
These are published results, not Glowing Spoon measurements. Benchmark percentages, F1 scores, and Elo ratings use different scales. Production evaluation also includes latency, cost, reliability, and human review effort on your workload.