Llama 4 Maverick employs a mixture-of-experts architecture, featuring 128 experts with a total of approximately 400 billion parameters, utilizing 17 billion active parameters per token. This design enables it to outperform models like GPT-4o and Gemini 2.0 in coding and reasoning benchmarks.
The model employs Meta’s mixture-of-experts (MoE) framework with active parameter counts of 400 billion. Maverick utilizes a large pool of 128 experts for task-specific activation.
Trained on curated datasets including multilingual corpora, image datasets, and synthetic reasoning examples.
Code Samples:
Detailed API Documentation is available here.
Llama 4 Maverick has implemented safeguards against misuse such as generating harmful content or violating user privacy during tool integrations.
Custom Llama 4 Community License
OpenAI-compatible — swap the base URL and it works with your existing SDK.
| Type | Price |
|---|---|
| Input | |
| Output | |
| Benchmark | Score | What it measures | Source | Retrieved |
|---|---|---|---|---|
| Intelligence | 9.3 | Composite score across standardised reasoning, knowledge and problem-solving evaluations, measured independently by Artificial Analysis | Source | September 12, 2026 |
| Coding | 16.3 | Composite score across standardised coding evaluations, measured independently by Artificial Analysis | Source | September 12, 2026 |
| Math | 19.3 | Composite score across standardised mathematics evaluations, measured independently by Artificial Analysis | Source | September 12, 2026 |
| Model | Input | Output | Context | Best for |
|---|---|---|---|---|
Llama 4 Maverick This page | ||||
| Reasoning + agents | ||||
| Balanced coding + agents | ||||
| Long-context, multimodal & agentic workflows | ||||
| Reasoning + agents |
Yes, Llama 4 Maverick can stream responses as they are generated.
Yes, Llama 4 Maverick supports both function calling and structured outputs.
Llama 4 Maverick was built by Meta.
Llama 4 Maverick is accessed through the v1/chat/completions endpoint.
