Alpha
Contact
Assess · Meta Llama 4

Can you own it?

Ownership levelPartialnone·limited·partial·substantial·fullAnalytical input C ยท 70.8/100

This page is a projection of the one entry record, the Use & modify and Transparency factors that Assess covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.

Which domain expands which factor
  • AssessUse & modify + Transparency
  • ImplementData control + Reliability
  • UseReliability
  • SupportTransparency

Intended & out-of-scope use

Llama 4 is Meta's natively multimodal mixture-of-experts generation: Scout and Maverick are released and Behemoth is the larger sibling. For non-EU users it is a capable, well-supported family for general, coding, multimodal and long-context work. It has its own entry, separate from the Llama 3.x text line (the meta-llama entry), because the licence differs.

It is out of scope for EU-domiciled entities. The multimodal licence is not granted to individuals domiciled in, or companies with a principal place of business in, the EU, so an EU-domiciled organisation cannot license these models. Only downstream end users of someone else's product are exempt.

Known limitations, bias & failure modes

The main limitation is legal. EU-domiciled entities have no licence to the multimodal models, the large models are systemic-risk, and the licence adds a 700M-MAU commercial trigger, "Built with Llama" naming and a download access gate. Training data and code are closed, and Maverick and Behemoth need very large infrastructure.

Openness tier & components

The tier is gated_open (dimension score 3 on access). Weights are downloadable after click-through acceptance and documentation is strong, but training data and code are closed and the licence is conditional. Licence conditions are scored under legal.

License terms & what you may do

The Llama 4 Community Licence is not OSI-approved. The incorporated AUP states: "with respect to any multimodal models included in Llama 4, the rights granted under Section 1(a) ... are not being granted to you if you are an individual domiciled in, or a company with a principal place of business in, the European Union". Downstream end users are exempt. Llama 4 is multimodal, so EU-domiciled entities get no licence.

The licence also keeps the 700M-MAU commercial trigger, "Built with Llama" naming and attribution, an AUP with five prohibited-use categories, and California governing law. The Llama 3.1 text models carry no EU carve-out. Legal therefore scores 2, and EU-domiciled ownership is effectively none.

Supply-chain provenance

The canonical source is the verified, access-gated meta-llama organisation on Hugging Face. It distributes safetensors with checksums and per-version licence tags and has no malicious-checkpoint incident. Bedrock, Vertex, Together and Groq host Llama 4. A hosted end-user product is the AUP-exempt path for EU end users, but not for EU-domiciled entities building on the weights. Pin the revision and verify checksums.

EU AI Act posture

Llama 4 is GPAI, and the large models are systemic-risk (>1e25 FLOP). The Article 53 open-source exemption does not apply, for two reasons: the Community Licence is not FOSS, and systemic-risk models are excluded from the exemption. Meta declined the EU GPAI Code of Practice and publishes no copyright policy or training-content summary.

The EU-multimodal carve-out denies EU-domiciled entities a licence outright. For EU-domiciled use, prefer the Llama 3.1 text models (the meta-llama entry) or a non-Llama family. Legal scores 2, with a hard flag.

Benchmarks & evaluation

Llama 4 Maverick is competitive among open multimodal models, and the family adds native multimodality and very long context. OneHill has not reproduced the benchmarks, so performance is capped at 4 and no specific figures are asserted as verified.

Independent safety evaluation

Meta ships Llama Guard (input/output moderation) and Prompt Guard (prompt-injection and jailbreak detection) with the family, alongside documented safety tuning. No broad independent multimodal red-team is available to aggregate, so safety scores 4, not 5.

How this scores

The ownership factors this domain covers, drawn from the one entry record.

1

Use and modify freelyCan you run, modify and adapt it with no gate and no field-of-use trap?

Moderate

Non-EU users may use and modify commercially below 700M MAU, subject to the AUP and "Built with Llama" naming, so moderate. For EU-domiciled individuals and companies the multimodal licence is not granted, so their use and modification is effectively none. Moderate is the global score; the EU carve-out applies on top for EU entities.

How this scores (AOI sub-dimensions)
Openness3/5how much is released - weights, data, code, licence - and how freelyWeights are downloadable after click-through acceptance and documentation is strong, but training data and code are closed and the licence is conditional.
Legal2/5how permissive and clean the licence is for real commercial useThe Community Licence is not OSI-approved and carries a 700M-MAU trigger and an AUP.
2

TransparencyDo you know what it is: weights, training, behaviour, and legible terms?

Moderate

Weights are inspectable and documentation is strong, but training data and code are closed. Open weights with a closed process score moderate.

How this scores (AOI sub-dimensions)
Provenance4/5how well we can trace and verify what went into the modelThe canonical meta-llama org on Hugging Face is verified and access-gated.
Governance3/5how accountable and well-documented the publisher isMeta is a named, accountable US provider with a verified org and a reporting path (LlamaUseReport@meta.com), which meets the score-3 anchor.
What this means for adoptionFor non-EU users you partially own self-hosted Llama 4. It runs on your infrastructure, performs strongly and pairs with Llama Guard and Prompt Guard, so reliability is strong. The 700M-MAU trigger, naming conditions and access gate keep use-and-modify at moderate, so ownership is partial. For EU-domiciled individuals and companies the licence grants nothing for the multimodal models, so effective ownership is none; prefer the Llama 3.1 text line (the meta-llama entry, no EU clause) or a non-Llama family. All users should check the 700M-MAU trigger, honour the Llama attribution terms and confirm domicile against the AUP.

Sources

The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.

Licenceread2026-11-19
The Llama 4 Community License (version effective 5 April 2025) sets a 700M-MAU threshold under Section 2, "Additional Commercial Terms" ("greater than 700 million monthly active users in the preceding calendar month, you must request a license from Meta").
Terms of serviceread2026-11-19
The Llama 4 Acceptable Use Policy states: "With respect to any multimodal models included in Llama 4, the rights granted under Section 1(a) ...
Model cardread2026-11-19
The verified, access-gated meta-llama Hugging Face org hosts Llama 4: Scout (109B/17B, very long context), Maverick (400B/17B) and Behemoth (larger), all natively multimodal MoE models in safetensors with checksums.
Third-party analysisunverified2026-08-03
On public leaderboards Llama 4 Maverick is competitive among open multimodal models, and the family adds native multimodality and very long context.
Model cardunverified2026-08-03
Meta ships Llama Guard (input/output moderation) and Prompt Guard (prompt-injection and jailbreak detection) with the Llama family and documents safety tuning.