# Llama 4 — weights, license and what is actually open

Meta’s first natively multimodal Llama generation and first Llama mixture-of-experts release.

Canonical: https://brightaifuture.com/open-models/llama
Format: model-family
Source publication: Not established
Bright publication: 2026-09-19
Substantive update: None recorded
Evidence and review: AI-assisted primary-source check on 2026-09-19; no independent model reproduction or license certification. This record does not independently reproduce Meta’s comparative benchmark claims.

## Documented release

Scout and Maverick

Organization: Meta

Llama 4 Scout and Maverick release: 2025-04-05 (day precision). Source: https://ai.meta.com/blog/llama-4-multimodal-intelligence/

Parameters: Scout: 109B / 17B active; Maverick: 400B / 17B active

Architecture: Early-fusion multimodal mixture-of-experts transformer.

Modalities: text, image

Context: Scout: 10M tokens; Maverick: 1M tokens

## License and commercial use

Llama 4 Community License Agreement

Commercial use is allowed subject to the custom license, including an additional term for services above 700 million monthly active users.

## What is open

Weights: available. Scout and Maverick weights are available under Meta’s custom terms. Source: https://ai.meta.com/blog/llama-4-multimodal-intelligence/

Architecture: available. Meta publishes architecture and model-card details. Source: https://ai.meta.com/blog/llama-4-multimodal-intelligence/

Inference code: available. Official downloads and ecosystem integrations support local inference. Source: https://ai.meta.com/blog/llama-4-multimodal-intelligence/

Training code: not-established. The complete code used to train Llama 4 was not released. Source: https://ai.meta.com/blog/llama-4-multimodal-intelligence/

Training recipe: partial. Meta describes key methods including multimodal pre-training and distillation. Source: https://ai.meta.com/blog/llama-4-multimodal-intelligence/

Data information: partial. The model card provides categories and governance information without the corpus. Source: https://ai.meta.com/blog/llama-4-multimodal-intelligence/

Training data: not-established. No sufficiently specific public artifact was confirmed in this review. Source: https://ai.meta.com/blog/llama-4-multimodal-intelligence/

Evaluation: partial. Meta reports benchmark results; full evaluation reproduction is not bundled. Source: https://ai.meta.com/blog/llama-4-multimodal-intelligence/

Commercial use: restricted. Commercial use is governed by a custom community license with scale-based conditions. Source: https://ai.meta.com/blog/llama-4-multimodal-intelligence/

## Uses and strengths described in sources

Native multimodality

Very long context on Scout

Large deployment ecosystem

## Hardware and quantization

Scout is described as fitting on one H100 in Int4; Maverick is described as fitting on one H100 host.

Meta states Scout can fit on one H100 with Int4 quantization

## Independent evidence

This record does not independently reproduce Meta’s comparative benchmark claims.

## Limitations

Custom license is not OSI-approved

Training data and full training stack are not released

## Provenance and history

{}

## Original sources

- [The Llama 4 herd](https://ai.meta.com/blog/llama-4-multimodal-intelligence/)

## Continue exploring

- [Open Models](https://brightaifuture.com/open-models)
- [Open intelligence thread](https://brightaifuture.com/threads/open)
- [Explaining a novice programmer's compiler error](https://brightaifuture.com/discoveries/open-model-pedagogy-compiler-errors)