Meta 发布 Llama 2 开放权重

第一代 LLaMA 的权重虽然在社区流传，正式条款却主要面向非商业研究。工程师已经会量化、微调和本地运行，企业仍很难把那条灰色路径写进采购与产品合同。模型能够下载，不代表公司能够放心上线。

2023 年 7 月 18 日，Meta 发布 Llama 2 的 7B、13B 与 70B 基础模型和 Chat 变体，宣布研究与商业用途免费可用。团队由此可以申请权重、在自己的环境中部署，也可以选择云厂商托管。微软是优先云合作方，但不是许可允许的唯一运行地点。

“免费可用”后面还有一整份 Llama 2 社区许可。它包含可接受使用政策、针对超大规模月活产品的额外授权要求，以及对某些用途的限制，例如用输出改进其他大型语言模型。这些条款使 Llama 2 与封闭 API 拉开距离，也使它不同于没有此类附加限制的 OSI 开源软件许可。开放权重、开放代码和开源许可，从这里开始更难被当成同义词。

模型本身也不只是第一代的重新授权。Meta 同时提供基础与对话微调版本，论文描述了更长训练、安全微调和红队测试。社区可以围绕同一批官方检查点继续做量化和领域微调，企业则第一次更容易把私有部署、数据边界和推理成本放进正式方案。技术评估与许可审查开始同时进行。这种中间位置扩大了选择，也带来新的工作。团队不再只能在封闭 API 与非商业研究权重之间二选一，却必须判断自己的用户规模、用途和分发方式是否落在许可范围内。权重放进私有环境，并不会免除内容安全、输出合规与模型局限；它只把更多控制和更多责任一起交给部署者。

Llama 2 最清楚的产品界面其实有两部分：一个让人把权重带走的下载入口，以及一份说明能怎样带走的许可文本。只看前者，会把“开放”读得过宽；只看后者，又会漏掉它确实为商业部署打开的空间。两者必须在同一页上读完。

Although first-generation LLaMA weights circulated through the community, the official terms were aimed mainly at noncommercial research. Engineers already knew how to quantize, fine-tune, and run the models locally. Companies still found it difficult to place that gray route inside procurement and product contracts. A model being downloadable did not mean a company could comfortably ship it.

On 18 July 2023, Meta released Llama 2 base and Chat variants at 7B, 13B, and 70B, free for research and commercial use. Teams could request the weights, deploy them in their own environments, or choose hosted cloud offerings. Microsoft was the preferred cloud partner, not the only location permitted by the license.

“Free to use” was followed by the full Llama 2 Community License. It included an acceptable-use policy, an additional authorization requirement for products above a very large monthly-active-user threshold, and restrictions on certain uses, including using outputs to improve other large language models. These terms separated Llama 2 from closed APIs and from OSI open-source software licenses without comparable conditions. Open weights, open code, and open-source licensing became harder to treat as synonyms.

The model release was more than a relicensing of the first generation. Meta provided base and chat-tuned versions, and the paper described longer training, safety fine-tuning, and red-team work. Communities could continue quantization and domain fine-tuning from the same official checkpoints. Enterprises could more readily place private deployment, data boundaries, and inference cost into a formal plan. Technical evaluation and license review now happened together. That middle position expanded choice and created work. Teams no longer faced only a binary between closed APIs and noncommercial research weights, but they had to determine whether their user scale, purpose, and distribution fell within the terms. Keeping weights in a private environment did not remove content-safety, output-compliance, or capability limits. It transferred more control and more responsibility to the deployer at the same time.

Llama 2's clearest product surface had two parts: a route for taking the weights away and a license explaining how they could be taken. Reading only the download button made “open” too broad. Reading only the restrictions missed the real space it created for commercial deployment. The release required both to be read on the same page.