llama-4-maverick

Llama 4 Maverick is Meta's advanced open-weights model in the Llama 4 generation, using a larger mixture-of-experts architecture than Scout — 400B total parameters with 17B active per token — delivering significantly higher capability while maintaining efficient inference costs. It supports native multimodal inputs, a 1M-token context window, tool use, and streaming. As an open-weights model, it can be self-hosted or accessed via third-party providers.

Llama 4 Maverick is available at approximately $0.20–$0.50 per million input tokens and $0.60–$1.50 per million output tokens from third-party providers, with Meta's own API offering free access subject to rate limits. At 1,000 input + 500 output tokens per call, it costs approximately $0.0005–$0.0013 per 1,000 API calls via third-party providers — significantly cheaper than proprietary frontier models at comparable capability levels. Self-hosted deployments require more substantial GPU infrastructure than Scout (typically 4–8× H100 80GB GPUs for efficient inference) but eliminate per-token costs entirely for high-volume workloads.

Llama 4 Maverick's 1M-token context window handles very large documents and codebases in a single call. Latency for self-hosted deployments on H100 clusters is approximately 300–600ms TTFT; third-party API providers vary. Throughput on H100 80GB is approximately 40–80 tokens per second depending on batch size and configuration. Maximum output is 16,384 tokens. The MoE architecture means that despite 400B total parameters, the active compute per token is closer to a 17B dense model — enabling more cost-effective inference than a dense model of equivalent capability.

Llama 4 Maverick is best suited to advanced workloads requiring strong reasoning, coding, and multimodal capability at lower cost than proprietary frontier models: complex code generation, document analysis, research synthesis, and agentic workflows. Its open-weights nature makes it suitable for applications with data privacy requirements (self-hosted, no data leaving your infrastructure), fine-tuning for domain-specific tasks, and deployments in regulated industries where proprietary API usage is restricted. Vision support handles images natively. For teams with GPU infrastructure and workloads that justify the hardware investment, Maverick delivers near-frontier capability at a fraction of proprietary API costs.

Llama 4 Maverick sits in the advanced tier of Meta's Llama 4 lineup, above Scout and below the unreleased Behemoth. Its main competitors are proprietary frontier models from OpenAI and Google (no self-hosting), and other open-weights models that are smaller and cheaper to self-host. For teams evaluating open-weights alternatives to proprietary frontier models, Maverick is the strongest option in the Llama 4 generation. Its intelligence score of approximately 85/100 reflects advanced-tier capability competitive with proprietary models at its price point. The open-weights licensing (Meta's custom license) permits commercial use subject to usage restrictions for very large deployments.