Llama 4 Maverick

Who should use this model?

• Server-side multimodal intelligence where stronger reasoning matters more than simple local install convenience or cost.

Choose something else when

• Avoid unless you have server-class GPUs, huge RAM, and can absorb license caveats comfortably today.

What hardware makes sense?

The download is only part of the memory needed. Leave room for the app, the input it processes, and other open programs. For chat models, context includes the instructions, conversation, and answer. Longer context needs more memory.

CPU only

Use a smaller model for this memory limit.

8GB system memory

Use a smaller model for this memory limit.

16GB system memory

Use a smaller model for this memory limit.

32GB system memory

Use a smaller model for this memory limit.

8GB GPU memory

Use a smaller model for this memory limit.

12GB GPU memory

Use a smaller model for this memory limit.

24GB GPU memory

Use a smaller model for this memory limit.

Apple unified memory

Use a smaller model for this memory limit.

Model facts that change the decision

Parameters

400B total / 17B active

Input

Multilingual text input · Image input · Text output · Code output

Published context limit

1,000,000 tokens

Model format

safetensors · GGUF · Ollama · MLX

Suggested compression

Q4_K_M for first local attempts unless a cited runtime recommends a different default.

Typical Q4 file

243GB before runtime and context overhead

Languages

Unknown

Using image input

This model needs an app or library that supports its image-processing components. Follow the exact model’s instructions for preparing image inputs. Some GGUF downloads also need a matching projector file; loading the text model alone does not establish image support.

Conditions that change the answer

• Maverick is officially released in BF16 and FP8; Meta says FP8 fits a single H100 DGX host, not a normal desktop GPU.
• The 1M-token context is not a beginner-local default; vLLM examples use 8x H100 or 8x H200 for long-context serving.
• Community GGUF and MLX conversions exist, but the smallest listed files are still very large and should be treated as experimental for local use.
• MoE lowers active compute but not stored-weight memory; the full 400B parameters still matter for RAM/VRAM.
• Llama 4 uses a custom Meta license, not Apache 2.0.
• Commercial use: Permitted for commercial and research use subject to Meta's custom Llama 4 Community License, Acceptable Use Policy, attribution requirements, 700M MAU threshold, and EU multimodal restriction.
• Llama 4 Maverick is the stronger Llama 4 model for reasoning, coding, and image work. It is not a normal local model; expect server-class memory, FP8 or aggressive quants, and careful licensing review before deployment.