Can you own it?
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.
- AssessUse & modify + Transparency
- ImplementData control + Reliability
- UseReliability
- SupportTransparency
Intended & out-of-scope use
Llama is a general-purpose assistant lineage for text generation, summarisation, retrieval-augmented question answering, coding help and multilingual chat. Meta positions the Instruct checkpoints for assistant use and the base checkpoints as starting points for further fine-tuning. Llama is the default self-hosted base for internal assistants, RAG systems and agentic prototypes.
Out of scope: any use forbidden by the Acceptable Use Policy (illegal activity, weapons, CSAM, targeted harassment, high-stakes unsupervised decision-making, and similar), and base checkpoints deployed directly as chat models. Base models are un-tuned and should never face users. Deploy the safety-tuned Instruct variant with Llama Guard instead.
Known limitations, bias & failure modes
Meta's model cards document the class-standard limitation profile. Llama hallucinates plausibly and confidently, can reflect training-data bias, and retains residual susceptibility to prompt injection and determined jailbreaks. Treat retrieved documents and tool outputs as untrusted input, never as instructions. Long-context recall degrades toward the far end of the window, and the smaller checkpoints trade capability for footprint. These limits are shared by current LLMs, which is why the safe-deployment controls are mandatory.
Openness tier & license classification
Llama sits in the gated_open tier, which caps the openness dimension at 3. Weights are
downloadable and the documentation (model cards, intended-use and limitations sections) is
strong. Access is gated behind a click-through acceptance on Hugging Face, both the training data
and the training code are closed, and evaluation is only partially released. The artifact is
documented but the recipe is not. The license classification is open_weight, not open source.
License terms & what you may do
The Llama Community License is not OSI-approved and should not be described as open source. It permits commercial use with conditions. Any deployer above 700 million monthly active users must obtain a separate license from Meta. An Acceptable Use Policy governs permitted applications, and "Built with Llama" attribution plus naming requirements attach to derivatives.
The text-only models in this entry carry no EU-domicile restriction (the multimodal carve-out is tracked in the separate llama-4 entry). Calling Llama "open source" is inaccurate on these terms; it is open-weight with conditions attached.
Supply-chain & provenance
The canonical source is the verified meta-llama organisation on Hugging Face. Distribution is
safetensors-only (no unsafe pickle load path from the canonical source), checksums are published,
and there is no malicious-checkpoint incident on Meta's own org, a checkpoint-checklist score of
roughly 6/8. Missing for full marks are cryptographic model signing (no Sigstore/OpenSSF
model-signing) and a SLSA-style provenance attestation.
The secondary ecosystem includes community GGUF quantizations (bartowski, unsloth) and hosted endpoints on Bedrock, Azure, Vertex, Together, and Groq. Each is a separate artifact whose trust equals its uploader or host. Pin the exact revision from the canonical org and verify the checksum before deploying.
EU AI Act posture
Llama is a GPAI model. The largest checkpoints (Llama 3.1 405B; the large Llama 4 models are in the llama-4 entry) exceed the 1e25-FLOP training-compute threshold, which triggers the systemic-risk presumption. Systemic-risk models receive no open-source exemption regardless of license, so the full Article 53 + 55 obligation set applies. The community license, with its acceptable-use policy, MAU threshold and access gate, likely fails the Act's "free and open-source" test in any case.
Meta's provider package is only partial against that bar. It has detailed model cards and safety evaluations (good downstream/Annex-XII information) but not a complete Article 55 systemic-risk package, and only partial copyright-policy and training-content-summary coverage. Meta publicly declined to sign the EU GPAI Code of Practice in 2025.
A downstream EU deployer receives an incomplete compliance file and must treat any multimodal Llama model separately. A deployer that fine-tunes and places a derivative on the EU market under its own name likely becomes a provider itself, with its own Annex XI/XII duties.
Benchmarks & evaluation
On public leaderboards Llama is consistently strong across reasoning, coding and instruction-following, and is a leader among open-weight generalist families. This is an aggregated, third-party read: performance is third-party framed, no specific numbers are asserted as verified, and the performance dimension is capped accordingly. Leaderboard standing is directional evidence of class-leading quality, not a reproduced measurement.
Independent safety / red-team results
Meta publishes its own safety evaluations in the model cards and ships two first-party safety classifiers as companion models: Llama Guard (content classification) and Prompt Guard (injection detection). No independent third-party red-team spans all harm domains, so coverage is marked partial. The publisher's safety evidence is useful but does not replace independent adversarial evaluation, the missing piece for a top safety score.
How this scores
The ownership factors this domain covers, drawn from the one entry record.
Use and modify freelyCan you run, modify and adapt it with no gate and no field-of-use trap?
ModerateOpen weights, but on Meta's terms. An access gate, the Community License's 'Built with Llama' naming and the 700M-MAU commercial-licence trigger are conditions on the grant. The EU-domicile restriction on multimodal models does not apply to these text-only 3.1 and 3.3 models (it sits in the Llama 3.2 and 3.3 AUPs and the Llama 4 licence; Llama 4 is the separate llama-4 entry).
TransparencyDo you know what it is: weights, training, behaviour, and legible terms?
ModerateYou can inspect and self-guard the weights, and Meta publishes safety evaluations plus Llama Guard and Prompt Guard. Training data and code are closed and no broad independent red-team spans every harm domain, so the finished model is visible while its training stays hidden.
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.