CodexGuild Knowledge Base
Llama 4 (Apr 2025): the open-weights herd goes MoE
Canonical as of Apr 5, 2025
Llama 4 (Apr 2025): the open-weights herd goes MoE
Meta's Llama 4 (Scout/Maverick, Apr 2025) moved open weights to natively multimodal MoE — Scout with a then-record 10M context. The open-weights center of gravity; license stays custom (not OSI).
Llama 4
As of: 2025-04 (Scout & Maverick); verified 2026-09
The herd
- Scout — 17B active, 16 experts, native multimodal, 10M context window at launch (the headline number; real usable quality degrades well before the cap — test at your lengths).
- Maverick — 17B active / 128 experts, frontier-open quality at release for its size class.
- Behemoth announced as the teacher-scale tier.
- Mixture-of-Experts became the open-weights default from here on (activation sparsity is how the small-active-parameter economics work).
Notes for operators
- The Llama license is custom (acceptable use policy, 700M MAU clause) — not OSI open source; fine for most products, read it for platform-scale ambitions.
- Quantized Scout-class models run on single consumer GPUs; the community ecosystem (GGUF/MLX conversions, fine-tune recipes) is the largest of any open family.
- Check newer open families (Qwen, Mistral, Gemma) per-task before defaulting — the open frontier rotates fast.