AI Potluck
Model components / Base / pretrained models

Llama 4 Scout / Maverick

Meta

Llama 4 herd, released April 2025: Scout (17B active / 109B total MoE, 10M-token context) and Maverick (17B active / 400B total MoE, 1M-token context), natively multimodal. Behemoth is still unreleased.

Consolidated on 2026-07-29 from llama-3-1; openness follows llama-4-scout-maverick, the release that currently governs. See docs/reference/identity.md. Reach and capability sit on different releases of the tier - the 3.1 8B checkpoint still supplies most of the downloads while Maverick supplies the capability reading. The meta-llama org lists no Behemoth repository. Verified 2026-08-13 via the Llama 4 Scout HF model card and the meta-llama organization page.

Openness

3 high confidence
3.0
weights
open(downloadable on HF)
data
closed
code
partial(inference/utils via llama-models, no training code)
license
Llama-4-Community(non-OSI: 700M MAU commercial cap as of Apr 2025 + no-compete/no-train-competing-LLM clause)

Weights are downloadable under Meta's Llama-4-Community terms: a 700M-MAU commercial cap, an acceptable-use policy and naming requirements. Llama 4 Scout and Maverick are the current release and Llama 3.1 carried equivalent terms, so nothing about the family's licensing has shifted. A license that bounds commercial use still permits it, and the scale places that a step above one that forbids commerce outright - CC-BY-NC or Mistral's Non-Production terms, which sit at 2. Meta's cap is a bound rather than a prohibition, which is why this lands at 3, the same place every other bounded-commercial license on the map lands.

  • https://huggingface.co/api/models/meta-llama/Llama-4-Scout-17B-16E recorded 2026-08-13

    `gated: manual`, `private: false`, safetensors weight shards listed, 18,815 downloads and 261 likes; `cardData.license` is `other`, so the Llama-4-Community tier rests on Meta's text rather than on a Hub label.

  • https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E recorded 2026-08-13

    Model card carries the LLAMA 4 COMMUNITY LICENSE AGREEMENT in full - "greater than 700 million monthly active users in the preceding calendar month, you must request a license from Meta" - and a Training Data section reading "~40 trillion tokens ... from a mix of publicly available, licensed data and information from Meta's products and services", which describes the corpus without releasing it.

  • https://api.github.com/repos/meta-llama/llama-models recorded 2026-08-13

    `private: false`, `archived: false`, description "Utilities intended for use with Llama models", license `NOASSERTION`. Utilities and inference helpers, no pretraining pipeline, which is what holds `code` at partial.

  • https://ai.meta.com/blog/llama-4-multimodal-intelligence/ recorded 2026-08-13

    Now answers HTTP 400 to a plain request; it was cited as a corroborating launch post and establishes no dimension, so nothing in the score rests on it.

  • https://huggingface.co/blog/llama4-release recorded 2026-08-15

    Hugging Face's Llama 4 release post, covering the Scout and Maverick open-weight checkpoints and their 256K context. Establishes that weights are published; the community-licence terms that cap the score are on Meta's licence text, not here.

  • https://royfactory.net/posts/ai/202512/meta-llama4-open-weights-scout-maverick-license/ recorded 2026-08-15

    Third-party summary of the Llama 4 open-weights release and its licence terms. Corroborates the open-weights reading; establishes no dimension on its own.

Adoption

4 medium confidence
4.0

1,335,865 downloads in the trailing 30 days across the five declared artifacts (meta-llama/Llama-3.1-405B 114,146; meta-llama/Llama-3.1-70B 41,103; meta-llama/Llama-3.1-8B 1,158,435; meta-llama/Llama-4-Scout-17B-16E 18,815; meta-llama/Llama-4-Maverick-17B-128E 3,366), which bands at level 4 (1M-10M) on the model adoption scale. The Llama 3.1 8B checkpoint still carries most of the family's reach even though the capability reading is taken on Llama 4 - reach and capability sit on different releases of the same family.

Capability

3 medium confidence
3.0

Llama 4 Maverick is the strongest released member of the family and so sets the score: Artificial Analysis puts it at 14 on the v4.1.1 Intelligence Index, ranked #26 of 44 and below the median of 18 for open-weight non-reasoning models of similar size. Llama 4 Scout scores 10 on the same index, and Behemoth was never released. Artificial Analysis has recalibrated the index, so numbers carried over from earlier versions of it are not comparable with these; on the current scale the family sits below its peer median, which is what a 3 records.

Verified 2026-08-13