Llama 4 Maverick
Meta·
LLMsopen-weightlocalMulti-GPU Self-Host
Overview
400B MoE (17B active) multimodal model. Matches Gemini 2.5 Pro and DeepSeek-V3 on reasoning benchmarks at significantly lower serving cost.
Capabilities and innovations
400B total / 17B active MoE128 experts, 1 activeNatively multimodal1M context windowOpen weights (Llama license)128-Expert MoE (Largest Open-Weight MoE)Early Fusion ArchitectureFrontier Performance at Low Active Params
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| GPQA Diamond | 69.8% | unsourced |
| MMLU | 92.4% | unsourced |
- Humanity’s Last Exam (no tools): not reported by the vendor
- SWE-bench Verified: not reported by the vendor
- SWE-bench Pro: benchmark did not exist at release
- MMLU-Pro: not reported by the vendor
- Terminal-Bench 2.x: benchmark did not exist at release
- MMMU-Pro: benchmark did not exist at release
Architecture and hardware
- Parameters
- 400B total, 17B active per token (MoE)
- Estimated VRAM at Q4
- ~230 GB, Multi-GPU Self-Host class
Links
More from Meta
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.