v2.2 · 2026-06-12

VISION / MULTIMODAL MODEL — RUNTIME REQUIREMENTS

·Commercial-use status from the supplied research: 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.

·Beginner summary from supplied research: 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.

Llama 4 Maverick

llama-4-maverick

MULTIMODAL

Llama 4 Maverick is a 400B multimodal model profile for local AI planning. This page records provider, license, context, quantization, hardware-fit estimates, setup hints, source links, and caveats so readers do not choose a model by name alone.

PROVIDER

Meta

FAMILY

Llama 4

MODEL TYPE

multimodal

PARAMETERS

400B (17B active)

MODALITIES

multilingual text input, image input, text output, code output

ARCHITECTURE

Mixture-of-experts autoregressive LLM with early-fusion multimodality; 128 experts

CONTEXT WINDOW

1,000,000 tokens

TRAINING TOKENS

—

RELEASE DATE

2025-04-05

LICENSE & LINKS

Llama 4 Community Licenselicense text

MODEL CARD

huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct

HUGGING FACE

meta-llama/Llama-4-Maverick-17B-128E-Instruct

SETUP HINTS

OLLAMA

ollama run llama4:maverick

LLAMA.CPP

Use a cited GGUF build with current llama.cpp-compatible tooling.

QUANTIZATION FILE SIZES (GGUF)

Q4_K_M

243 GB

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

RAM / VRAM ESTIMATES

MIN RAM

256 GB

COMFORTABLE RAM

512 GB

These are conservative local-inference estimates, not official hardware requirements.

HARDWARE FIT

CPU only (no GPU)

8 GB RAM

16 GB RAM

32 GB RAM

8 GB VRAM

12 GB VRAM

24 GB VRAM

Apple Silicon (unified memory)

Hardware fit values are conservative local-inference estimates based on GGUF size plus runtime and KV-cache overhead. Actual requirements depend on context length, quantization, runtime, GPU offload, and other running apps.

BEST FOR

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

AVOID IF

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

CAVEATS

·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.

SOURCE URLS