Tier 2 Restricted open-weight
What the publisher claims: open-source, open-weight
The one-liner: Influential open-weight family with custom conditions, now partly superseded inside its own publisher.
The six questions
Llama 4 Scout and Maverick weights remain downloadable from the meta-llama organization on Hugging Face and from Meta's own download site, gated behind license acceptance rather than closed; Meta's successor Muse weights are public as well.
The Llama 4 Community License is not an OSI-approved grant and imposes a 700M MAU separate-license trigger resolved at Meta's discretion, attribution and naming conditions, an EU-based exclusion from multimodal rights, and an Acceptable Use Policy; the Muse Model License adds a revenue-triggered agreement for MaaS providers.
Deployable on chosen hardware with standard runtimes, but subject to the license conditions, the gating on download, and the naming and attribution obligations that travel with redistribution.
The Llama 4 model card describes data categories, scale, cutoff, and environmental information, and the successor Muse releases publish model cards, but no reproducible corpus or recipe is available.
Limited: the training system stays closed, and Meta's shift from Llama to the proprietary Muse Spark flagship in April 2026 shows how quickly even the most influential open-weight line can move toward closed.
Meta's llama-models and Muse repositories take the occasional outside pull request and the Hugging Face org is publicly readable, but no Llama or Muse training run, dataset, or release decision is open to non-Meta contributors.
License
Llama 4 Community License plus Acceptable Use Policy for Scout and Maverick; Muse Model License for Muse Spark Open; Apache 2.0 for Muse Glimmer 30B — license source ↗
- Products above 700M monthly active users must request a separate Meta license, granted at Meta's sole discretion
- Mandatory 'Built with Llama' attribution and a 'Llama' naming prefix for derivatives
- Multimodal rights are withheld from individuals and companies based in the European Union
- The Muse Model License requires a separate agreement for Model-as-a-Service providers above $20M gross revenue in any 12 months
- Llama 4 repositories are gated on Hugging Face behind license acceptance
What is actually released
- weights (Llama 4 Scout and Maverick, gated on the meta-llama organization)
- model card describing data categories, scale, cutoff, and environmental information
- Muse Glimmer 30B weights under Apache 2.0
- Muse Spark Open weights under the Muse Model License
Why it matters
Llama is the essential case study in the politics of naming: it made open weights mainstream, its custom terms show why downloadable does not mean unrestricted, and its April 2026 retirement followed by an Apache 2.0 return under the Muse name in August 2026 shows the terms can change with the business strategy.
Sources
- https://dev.meta.ai/llama/llama4/license
- https://huggingface.co/meta-llama
- https://venturebeat.com/technology/meta-returns-to-open-source-with-muse-glimmer-an-apache-2-0-licensed-30b-parameter-ai-model-optimized-for-agents-available-now
- https://moelueker.com/blog/how-to-use-llama4-for-free
- https://www.deepset.ai/blog/open-source-llm-open-weight-models
Derived from the openness dataset behind But Is It Open? — snapshot 2026-11-07 ↗. Scores are qualitative readings of the cited evidence, not verdicts; tier measures openness, not intelligence.