Ethical Stack

Llama 4 family

Open-weight multimodal models under Meta's Community License, including Llama 4 Scout and Maverick. The weights are available, while terms and scale restrictions mean this is not unrestricted open-source software.

Score calculation

(4 + 4 + 7 + 6 + 3) ÷ 5 = 4.8

5 of six dimensions rated. Unrated dimensions and BDS status are excluded.

Full scoring method ↗
Setup details

Llama model release/license statements are model-family-specific. Meta platform data practices and public human-feedback descriptions are company/family context and not proof of every Llama 4 training input.

Meta's Llama documentation still lists Llama 4 Scout and Maverick as its newest Llama releases on 2026-10-23. Meta separately markets Muse models; those are outside the requested Llama family record.

Energy

↑

Meta's Llama 4 model card reports pretraining at 7.38M H100 GPU-hours (Scout 5.0M, Maverick 2.38M), 700 W peak capacity per device, and estimated 1,999 tonnes CO2e location-based emissions. It reports 0 market-based tonnes based on Meta's company-wide clean-electricity matching. These are Meta's reported estimates; GPU-hours and rated device power are not measured facility electricity. The figures exclude fine-tuning and do not report inference energy.

Reasoning & sources

Meta’s Llama 4 card reports GPU-hours and location-based CO2e for pretraining, but excludes fine-tuning and does not measure inference; its zero market-based figure relies on company-wide clean-electricity matching. The disclosed estimate is useful but incomplete, so compare the location-based boundary and ask for downstream serving impacts.

A high-impact pretraining disclosure is unusually specific and renewable matching is reported, but location-based emissions are estimated at 1,999 tonnes and inference/post-training footprints remain undisclosed; score reflects both scale and reporting transparency.

No measured facility kWh/PUE, water, post-training footprint, or standardized inference energy per task/token. The 0 market-based CO2e relies on Meta-wide renewable matching, not a model-specific zero-emissions measurement.

Training rights

↑

Meta's Llama 4 release describes its training approach but does not provide a full source-level dataset/rightsholder manifest in the checked material. Meta's general AI disclosures say it uses public, licensed, and Meta-product information; those policies do not establish legal rights for each Llama 4 work.

Reasoning & sources

The release plus Meta’s filing describe data practices and rights processes, but do not provide a complete Llama 4 source/rightsholder manifest. These policies provide some process disclosure, but do not establish that every training item was licensed or consented.

General data categories are disclosed but Llama 4 has no complete source-level rights inventory.

No complete Llama 4 dataset list, rights audit, or independent consent review found.

Privacy

↑

Meta says public posts and interactions with its AI products can be used to develop AI; it says private messages with friends and family are not used unless shared with AI. This is platform-level policy context, not inference privacy of self-hosted Llama weights.

Reasoning & sources

Meta’s filing distinguishes public interactions and AI interactions from private messages, which are generally not used unless a user shares them with AI. This is a policy boundary rather than a guarantee across every product or region; users should review settings and avoid sharing private material into AI features.

Weights can be self-hosted for operator-controlled prompt handling; this is a deployment option rather than a model guarantee.

Local deployment shifts prompt handling to the operator; Meta's platform policy is not applicable to every downstream host.

Openness

↑

Weights and model cards are downloadable, but access is governed by the Llama 4 Community License Agreement, including conditions and scale-related provisions. Training data and full end-to-end training pipeline are not fully released with the weights.

Reasoning & sources

The downloadable weights and model card enable local inspection, but the community license carries conditions and the training corpus is not fully open. The family is more accessible than a hosted-only model, while restrictions and missing data limit full reproducibility.

Weights are available, but a custom community license and incomplete training-data release limit openness.

Review the specific license and acceptable-use terms before treating a use as permitted; no full pretraining corpus is provided.

Worker treatment

↑

Meta has described human feedback and human annotations for earlier Llama generations. The sources checked do not disclose Llama 4 annotator workforce conditions, compensation, vendor chain, or worker protections.

Reasoning & sources

Meta’s earlier Llama responsibility materials discuss human feedback and annotation, but the Llama 4 announcement does not disclose current workers’ pay, vendors, or conditions. Those older-generation sources do not establish Llama 4 labor practices; current supplier details are needed.

Earlier-generation human-feedback descriptions cannot establish Llama 4 labour conditions.

Government & military

↑

Meta announced in September 2025 that Llama models were being used by US military and intelligence stakeholders, including a Legion Intelligence collaboration with USSOCOM on SOFChat; it also named Lockheed Martin and Anduril among defense partners and described Llama use in military workflows. These are Llama-family/company disclosures and do not identify Llama 4 as the exact deployed checkpoint.

Reasoning & sources

Meta’s 2025 announcements describe Llama use by military/intelligence stakeholders and named defense partners, including a USSOCOM-linked tool, while not identifying the exact checkpoint. That is a documented family-level relationship and a serious concern; a buyer should verify whether Llama 4 is used in its military workflow.

Meta publicly describes real military and intelligence use and named defense-industry partners at the Llama-family level; exact Llama 4 deployment remains unverified.

The announcements are dated 2025 and do not give a full deployment inventory or identify exact model versions for each named use.

BDS status

↑

Meta is named as a pressure target in the BNC’s official campaign sources. The product is developed or supplied by Meta.

Reasoning & sources

If following this BNC campaign is a requirement, consider alternatives to paid Meta products where practical. Running downloaded weights locally is a different decision from paying the developer or a cloud host; check the host as well. This label follows the named company’s campaign status; it does not claim that BDS separately reviewed this exact product.

This review covers the named targets in the BNC consumer/pressure guide and No Tech campaign, not every BDS campaign, local group, investment or supplier. Lists and ownership can change. A company’s absence from these lists is not an endorsement or proof that it has no relevant relationships.

All sources

Side by side