Llama 4

AI Model Multimodal Prior generation

Meta's MoE generation — and the stumble that triggered the superintelligence pivot.

You're looking at an AI model record: who built it, what it can do, and the hardware and data behind it.

Llama 4 (April 2025) moved to mixture-of-experts with native multimodality: Scout (109B/17B active, 10M-token context claim) and Maverick (400B/17B active), with the 2-trillion-parameter Behemoth teacher announced but never released. A mixed reception — benchmark controversy, delayed flagship — preceded Zuckerberg's Superintelligence Labs reorganization and hiring spree. In April 2026 MSL replaced the Llama line with Muse Spark — launched, in a historic reversal, without open weights.

How this connects to the rest of the stack

This record's links at a glance — click a group to jump to it.

Developed by

Powers

Meta AI

Trained on

NVIDIA H100

Successor to

Llama 3.1 405B

Competes with

Qwen3 DeepSeek-V3 Gemma 3

Sources

Llama 4 ↗

Record reviewed as of 2026-08-31 — see Methodology.