MDL-3468 EST.2025 · IDX.740 Language model In production ⇄ Compare
Llama 4 Meta AI · USA · 2025
Meta's natively multimodal open-weights family — Scout, Maverick and the giant Behemoth.
Llama 4 brought native multimodality and mixture-of-experts efficiency to Meta's open-weights line, with Scout offering an industry-leading context window of up to 10 million tokens. Powering Meta AI across WhatsApp, Instagram and Facebook, the Llama ecosystem surpassed a billion downloads and anchors much of open-source AI.
#open weights #multimodal #MoE #long context
Official website ↗ ▸ SpecificationsClass & identity
Model class Open-weights multimodal LLM
Version / release 2025 Architecture
Architecture Mixture-of-Experts, native multimodal
Parameters Scout / Maverick / Behemoth
Context window Up to 10M tokens (Scout)
Modalities Text, image → text
Knowledge cutoff / data — Performance (benchmarks)
MMLU —
GPQA Diamond —
SWE-bench —
MATH / AIME —
Chatbot Arena Elo —
Domain benchmark — Capabilities Text Vision Long context (≥200k) Open weights Fine-tunable Multilingual Function calling
Reasoning mode —
Tool use / agentic Function calling
Multilinguality Multilingual
Key strengths Ultra-long context, open ecosystem (1B+ downloads) Access & deployment
Availability Open weights
License Llama Community License
Pricing No vendor price. Meta publishes the weights under the royalty-free Llama 4 Community Licence and sells no hosting of its own, so the cost is whatever a third-party host charges. Above 700M monthly users a separate licence is required.
Deployment Self-host / cloud / product-embedded
Safety / alignment Llama Guard tooling Compiled to the wujec.ai Spec Standard v1 — “—” means not publicly disclosed.