Llama 4 Maverick

The Llama 4 Maverick is a versatile model in the next generation of the Llama 4 family, released in April 2025. Unlike the more compact Scout, the Maverick is designed for those who need to maximize performance and processing power. The model uses MoE architecture with 128 experts, which has 400 billion parameters in total - making it one of the largest scale open source models on the market. However, only 17 billion parameters are active at each step, significantly reducing inference latency and resource requirements. Like all models in the Maverik series, Maverik supports native multimodality, making it easy to input and process images together with image support text.

With its large-scale architecture, Maverick offers consistency and depth of reasoning, accuracy and security of responses. The model performs respectably on complex logical reasoning, math, and programming tasks, outperforming previous generations of LLMs and major competitors on several key metrics.

The model incorporates advanced security features such as Llama Guard 4 (12B) to filter unsafe content and Llama Prompt Guard 2 (22M and 86M) to protect against manipulative or malicious input.

Llama 4 Maverick is ideal for complex enterprise solutions and large-scale projects where deep insight and precise execution are required.

Announce Date: 05.04.2025
Parameters: 402B
Experts: 128
Activated at inference: 17B
Context: 1M
Layers: 48, using full attention: 12
Attention Type: Chunked Attention
VRAM requirements: 268.9 GB using 4 bits quantization
Developer: Meta AI
Transformers Version: 4.51.0
License: LLAMA 4

Use our pre-built public endpoints for free to test inference and explore Llama 4 Maverick capabilities. You can obtain an API access token on the token management page after registration and verification.

Rent your own physically dedicated instance with hourly or long-term monthly billing.

We recommend deploying private instances in the following scenarios:

• maximize endpoint performance,
• enable full context for long sequences,
• ensure top-tier security for data processing in an isolated, dedicated environment,
• use custom weights, such as fine-tuned models or LoRA adapters.

Contact our dedicated neural networks support team at nn@immers.cloud or send your request to the sales department at sale@immers.cloud.