Llama 4 Maverick pricing: Meta's open multimodal model with 1M context

Llama 4 Maverick is an open-weight multimodal model from Meta with vision support, a 1M-token context window, and a mixture-of-experts architecture. Self-hosting has no per-token API fee, but it still carries license obligations, hardware, energy, security, and operations costs. Sampled third-party inference routes list input around $0.20/1M tokens.

By benchr Editorial Team · · Provider figures checked against cited sources; memory lower bounds are parameter-count arithmetic · View changelog

License Costopen weights
Input / 1Mvia Together.ai
Contextmax window
Visionimage input

Pricing options

llama-4-maverick — pricing options
OptionCost
Self-hosted (Meta weights)Infrastructure only
Together.ai input~$0.20/1M
Together.ai output~$0.70/1M
Context window1,000,000 tokens

The case for open-weight multimodal

Most managed multimodal APIs send image data to provider infrastructure, which creates data-residency and privacy considerations. A correctly isolated Maverick deployment can keep inference data inside an approved environment, but local weights alone do not guarantee that outcome: telemetry, remote tools, logging, storage, monitoring exports, and administrator access still need review. Compare the deployment controls and task quality on the exact image workload before choosing it.

1M context: where it matters

A 1M token context window handles approximately 750 pages of text or an equivalent combination of text and images. For an open-weight model, this context length is unusual, although supported limits vary across open-weight deployments. Candidate uses include large codebase slices, multi-document processing, and long agent sessions. The published ceiling is not a guarantee of retrieval quality or memory fit at the full length.

Hardware and deployment considerations

Maverick has 400B total parameters. Weight storage alone is roughly 800GB at BF16 (400B × 2 bytes) and 200GB at ideal 4-bit packing, before KV cache, runtime buffers, quantization metadata, replicas, and serving headroom. Its MoE design reduces active compute per token, not the memory needed to make the expert weights available. Plan a sharded or offloaded deployment from the exact checkpoint, precision, context target, concurrency, and serving stack; do not infer fit from active-parameter count alone.

Cost scenarios

Via Together.ai at 10M input + 3M output per month: approximately $2 + $2.10 = ~$4.10/month. Gemini 3.1 Pro (multimodal, 1M context) at the same volume: $20 + $36 = $56/month — 14× more expensive.

The lower rate matters only if Maverick clears the workload's quality threshold. Its published vision results trail frontier closed models on some complex image-reasoning tasks, so compare accepted-result cost rather than token price alone.

Use-case fit

Consider for: Self-hosted multimodal pipelines where image data residency matters; long-context applications that need up to 1M tokens; teams with existing GPU infrastructure; and bounded image tasks such as OCR, document classification, or product-image recognition.

Skip if: Your vision tasks require frontier reasoning quality — complex scene understanding, medical image analysis requiring high accuracy, or sophisticated visual QA where closed-model performance gaps are measurable. Also skip if you need video or audio (image-only).

Decision checklist

For vision tasks: test Llama 4 Maverick against your specific image understanding requirements. OCR and structured document extraction are strong; complex scene reasoning and fine-grained visual detail tasks may show gaps versus Gemini 3.1 Pro or GPT-5.5.

Review the Meta Llama license terms for your deployment scale. The 700M MAU threshold is not a concern for most teams, but the "cannot train competing LLMs" restriction may affect AI product companies.

Frequently asked

How does Llama 4 Maverick compare to Llama 4 Scout?

Maverick is larger and more capable (reasoning, vision quality, 1M context). Scout is lighter, faster, and offers 10M context — the largest available window for extreme long-context retrieval. Choose Maverick for task quality; Scout for 10M context or maximum throughput on a budget.

What are the hardware requirements for self-hosting Llama 4 Maverick?

Maverick has 400B total parameters. Weight storage alone is roughly 800GB at BF16 and 200GB at ideal 4-bit packing, before KV cache, runtime buffers, quantization metadata, and serving headroom. MoE lowers active compute per token; it does not make the inactive expert weights disappear from memory. Size a sharded or offloaded deployment from the exact checkpoint, precision, context target, concurrency, and serving stack rather than a fixed GPU count.

Does Llama 4 Maverick have any commercial use restrictions?

Meta Llama Community License: commercial use allowed, attribution required, cannot train competing general LLMs, 700M+ MAU requires separate license. Not a barrier for most applications. Verify current terms at llama.meta.com/llama4.

Changelog

  • — Replaced undersized fixed-GPU claims with weight-memory lower bounds and separated MoE active compute from total model storage. Qualified local data-control language.
  • — Expanded with multimodal analysis, hardware requirements, license guide, and cost scenarios.
  • — Published.

Sources