Alpha
Contact
Support · Meta Llama 3.x

Will it last?

Ownership levelPartialnone·limited·partial·substantial·fullAnalytical input B ยท 72.4/100

This page is a projection of the one entry record, the Transparency factor that Support 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

Common problems & fixes

The two most frequent problems are operational, not model quality.

Gated-access failures (401/403, "you do not have access") occur when the license has not been accepted on the model page or the Hugging Face token has not been granted access. Approval is not always instant. Use the same account for acceptance and token.

Tokenizer and chat-template mismatches produce rambling, role-confused, or never-terminating output. This happens with hand-built prompt strings or the wrong special tokens. Render prompts with the tokenizer's apply_chat_template and configure the end-of-turn token as a stop token in your serving stack. Quantized community builds occasionally ship a stale tokenizer/template, so cross-check against the canonical meta-llama repo.

Versions & changelog

Meta ships on a predictable generational cadence: Llama 3.1 (8B/70B/405B), Llama 3.3 (a 70B refresh narrowing the gap to 405B), then Llama 4 (Scout/Maverick/Behemoth, mixture-of-experts, multimodal). Each generation is released with model cards documenting the changes. There is no rolling per-commit changelog for the weights. The durable way to track and pin a build is the Hugging Face revision (commit hash) on the canonical org. Pin a revision in production instead of tracking main.

Security disclosure process

Meta provides a Responsible Use reporting path for model issues, policy violations, and feedback (via the Responsible Use Guide and associated reporting channels). There is no formal CVE-style vulnerability-disclosure process and no cryptographic signing for the model weights, which is the gap noted in the governance score. Coverage is partial: model behaviour can be reported, but no security-response SLA or signed-artifact verification path is published.

Community & support channels

Support is community- and platform-based, with no formal support contract. Channels are the Meta Llama community resources and documentation on llama.com, Hugging Face model discussions on each repo, and Meta's GitHub repositories (llama-models, llama-recipes/cookbook, torchtune) for issues and examples. The serving-ecosystem projects (vLLM, llama.cpp, Ollama) also carry active Llama-specific issue trackers.

Tracked known issues

The model cards document the standing known issues: hallucination in the class-standard profile, residual jailbreak and prompt-injection susceptibility, and the explicit warning that base checkpoints are not safety-tuned and must not be deployed as chat models. Long-context recall softens mid-window, and multilingual quality outside the eight officially supported languages is not validated. These caveats persist across releases, and the safe-deployment controls are designed to contain them.

How this scores

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

2

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

Moderate

You 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.

How this scores (AOI sub-dimensions)
Provenance4/5how well we can trace and verify what went into the modelVerified Meta org on Hugging Face, safetensors-only distribution, published checksums, clear canonical source, and no malicious-checkpoint incident on the canonical org (checklist ~6/8).
Governance4/5how accountable and well-documented the publisher isReputable, legally accountable publisher with a predictable multi-generation release cadence, a Responsible Use Guide and a reporting path.
What this means for adoptionYou own the self-hosted Llama weights: they run on your infrastructure, perform at the top of the open field, and pair with Meta's Llama Guard and Prompt Guard. Ownership is only partial. The Community License sets conditions on the grant: an access gate, 'Built with Llama' attribution and a 700M-MAU commercial trigger. You cannot freely use and modify these models the way an OSI licence allows, and training is closed. Check the MAU trigger and honour the attribution terms first.

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
Llama 4 Community License.
Licenceread2026-11-19
Llama 3.1 (release date 23 July 2024) and Llama 3.3 (release date 6 December 2024) Community Licences.
Terms of serviceread2026-11-19
Llama 4 Acceptable Use Policy: five prohibited-use categories (illegality/rights infringement; death or bodily harm incl.
Third-party analysisread2026-07-25
Llama family variants, EU multimodal licensing restriction, ecosystem/cloud availability, and Meta declining the EU GPAI Code of Practice.
Model cardread2026-11-19
Canonical, verified and gated meta-llama organisation on Hugging Face distributing safetensors with checksums; licence tags license:llama3.3 and license:llama3.1 on the 3.x repos (Llama 4 reads license other with license_name llama4).
Third-party analysisread2026-07-25
Llama is competitive-to-leading among open-weight generalist families on public leaderboards.
Documentationunverified2026-07-25
Meta ships safety-tuned instruct variants plus Llama Guard and Prompt Guard companion classifiers with published safety evaluations and a responsible-use reporting path.
Model cardread2026-07-25
Llama 3.1 Instruct model card on Hugging Face: 128K context window, the Llama 3 chat template and special tokens, the eight officially supported languages, and documented tool-calling.