← Models

Llama 4 Maverick

Meta · Llama 4

Open-weights Llama 4 tier for teams needing customization, control, and self-hosting flexibility.

Part of Llama family · Other versions: Llama 4 Scout
Type
multimodal
Context
262K tokens
Max Output
33K tokens
Status
current
API Access
Yes
License
Llama Community
open-weights self-hosted reasoning enterprise customization
Released April 2025 · Updated March 6, 2026

Overview

Freshness note: Model capabilities, deployment options, and licensing terms can change. This profile is a point-in-time snapshot last verified on February 15, 2026.

Llama 4 Maverick is Meta’s larger open-weight Llama 4 release for teams that prioritize deployment control and model customization. Meta’s official launch materials describe Maverick as a natively multimodal MoE model with 17B active parameters and 128 experts, designed to deliver frontier-style quality with self-hosting flexibility.

Capabilities

Maverick-class open models are often used for internal assistants, controlled-domain reasoning, multimodal workflows, and custom tool-enabled systems. They are especially useful when organizations need deeper tuning or policy-constrained deployment patterns.

Technical Details

Meta’s official Llama 4 launch materials position Maverick as the higher-capability sibling to Scout, built to run on a single H100 host. Real-world quality still depends heavily on serving stack, quantization choices, and evaluation discipline.

Pricing & Access

Access can come through self-hosted infrastructure or cloud providers exposing compatible endpoints. Cost structure differs substantially from closed APIs because infrastructure and operations become major factors.

Best Use Cases

Good fit for regulated environments, on-prem or private-cloud assistants, and teams that want deeper control over model lifecycle and deployment economics.

Comparisons

Compared with GPT-5.4 and Claude Opus 4.6, Maverick offers more deployment control but usually requires more engineering effort for equivalent polish. Compared with Qwen3-Max, choice depends on licensing, language needs, and serving strategy.