FrankX.AI
Meta AIGA

Llama 4 Maverick

Meta’s open flagship by default — permissively licensed and multimodal, but no longer leading the open pack.

Read the full Llama 4 Maverick analysis

Context

1.0M

Max output

16K

Input /1M

$0.20

Output /1M

$0.70

Live pricing via OpenRouter

Best for

  • Permissively-licensed self-hosting in your own VPC
  • Native open-weight multimodal (text+image) work
  • Long-context reasoning via the 10M-token Scout sibling

Watch out

A data-center model (8x H100 for Maverick FP8), not a consumer-GPU one, and its benchmarks now trail DeepSeek V4, Qwen3.7-Max and Kimi K2.6. The headline 1417 LMArena Elo was an experimental variant, not the public weights.

For creators. Self-host Scout on a single H100 for whole-corpus (10M context) multimodal drafting where you need data control and zero per-token cost.

Benchmarks

mmlu pro80.5
gpqa diamond69.8
livecodebench43.4

Capabilities

  • Native text+image multimodal input via early fusion
  • 1M-token context (Maverick); Scout sibling offers 10M
  • Llama 4 Community License — free commercial use under 700M MAU
  • Self-host: FP8 weights need an 8x H100 80GB node (~600GB); Scout (109B/17B) self-hosts on one H100 at ~55GB Q4
  • Run via vLLM / Ollama (Scout) / HF, or hosted APIs (~$0.15/$0.60 per 1M)

Sources