Availability changed the deployment choice
Llama 2 made commercial use available under a community license in July 2023, with attribution and acceptable-use obligations, a restriction on improving other large language models, and additional terms for very large services. DeepSeek described R1 as MIT-licensed in January 2025. Those are materially different terms; “downloadable” does not mean unrestricted.
The distilled model has its own provenance
DeepSeek’s release repository distinguishes the main R1 model from smaller distilled models built on Qwen and Llama bases. Its January 2025 documentation identifies Apache 2.0 origins for the Qwen variants and the relevant Llama licenses for the Llama variants. A family name alone is not enough to identify the artifact or its terms.
Weights are only part of reproducibility
OSI’s Open Source AI Definition 1.0 asks for data information, training and inference code, and parameters, alongside freedoms to use, study, modify, and share. The R1 paper explains a training approach, while the model release supplies weights. Neither fact by itself establishes that every artifact needed to reproduce training is available.
Evaluate control, not just access
Separate three decisions: whether the license permits the intended use, whether the deployment can meet cost and reliability requirements, and whether the evidence supports the task. Self-hosting can give an operator more control over where inference runs, while also making the operator responsible for serving and evaluation. Compare the specific model and its terms, not an entire family under one label.
How the story developed
Llama 2 opens a commercial route
Meta makes a downloadable model family available under its community license.
Open event record →R1 and its technical report are released
DeepSeek announces the model, distilled variants, and a technical report hosted in its repository.
Open event record →The paper receives an arXiv record
The versioned arXiv submission provides a stable citation distinct from the release date.
What remains open
- Which license applies to the exact base or distilled model being evaluated?
- What does serving cost at the required latency and utilization?
- Which release claims have been reproduced independently on the intended workload?
Newer developments to investigate
Unreviewed research leads after this brief’s evidence cutoff.
Matched “open weights” in the title.
Matched “open weights” in the summary.
Matched “open weights” in the title.
Matched “open weights” in the summary.
Matched “open weights” in the title.
Source notebook
What each source supports, and how to read it.
Meta and Microsoft Introduce the Next Generation of Llama ↗
Establishes commercial availability; consult the linked community license for its conditions.
Release and licensing statements from the provider, not independent performance evidence.
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning ↗
Describes the training approach; benchmark results are reported by the model’s developers.
LLAMA 2 COMMUNITY LICENSE AGREEMENT ↗
Version-specific terms: redistribution, acceptable use, competing-model restriction, and additional commercial terms.
The Open Source AI Definition – 1.0 ↗
OSI’s criteria distinguish parameters, code, and data information; a definition, not a model certification.
Pinned repository revision records base-model provenance and the distinct Llama and Qwen distill families.
Brief history
2026-09-05 · Initial foundations brief distinguishes weight access, license terms, and reproducibility; newer related records remain separate research leads.
2026-09-05 · Added the Llama 2 license, OSI definition, and a pinned R1 model card. Distinguished the release from its arXiv posting and clarified distilled-model provenance.