Meta 发布 Llama 3

2024 年 4 月 18 日，Meta 发布了 Llama 3，首批两个版本：8B 和 70B。如果只看名字，这像是一次常规迭代；但发布后的反响说明它远不止于此。在众多基准测试上，Llama 3 的表现直逼当时闭源的顶级模型——这是开源模型第一次真正站到了前沿的门口。

Llama 3 的底气来自 Meta 在数据与工程上的重注。它使用高达 15T token 的训练数据，规模远超前代；训练策略强调数据质量与后训练对齐。发布时还配备了改进的许可证与更友好的商用条款。这些细节让它不只是「又一个开源模型」，而是把「开放权重+前沿性能+可商用」三者同时做齐的产品。

社区反应异常热烈。研究者拿它做基座微调，创业公司拿它做私有化部署，个人开发者用它的 8B 版本在本地跑模型。Llama 3 几乎成为开源模型的事实标准，海量的适配、量化、评测项目随之涌现。一个更现实的变化出现了：企业不再只能依赖闭源 API，开源模型成了真正可选的替代。

7 月，Meta 又发布了 405B 的 Llama 3 旗舰，并配套推出完整的工具生态。405B 在多项能力上追平甚至超过当时的闭源竞品，把「开源模型到达前沿」从一次偶然变成了一贯事实。与 2023 年的 Llama 2 相比，Llama 3 完成了一次质的跨越——它不再是被拿来「追赶」的对象，而是定义基准的参与者。

Llama 3 的深远影响，是让「开源 vs 闭源」成为整个行业必须认真回答的问题。闭源厂商开始公开回应开源模型的挑战，政府与学界围绕开放权重展开了持久的政策辩论。对普通开发者来说，最直接的变化是选择变多了：模型不再是几家公司的专利，而是可以下载、修改、部署在自己服务器上的东西。

回看 2024 年 4 月，Llama 3 的意义不在于某一次得分，而在于它改写了竞争的基本面。它证明开源模型有能力到达前沿，这个事实改变了整个 AI 生态的走向——后来的 DeepSeek、Qwen 等国产开源模型的崛起，都与 Llama 3 打开的局面一脉相承。开源不再是「退而求其次」，而是与闭源并行的主流路线之一。

On April 18, 2024 Meta released Llama 3, starting with two versions: 8B and 70B. Judged by name alone it looked like a routine iteration; the reaction showed otherwise. On many benchmarks Llama 3's results pressed right up against the top closed models of the time—the first time an open model truly stood at the frontier's door.

Llama 3's confidence came from Meta's heavy bets on data and engineering. It trained on up to 15 trillion tokens, far beyond its predecessor, with an emphasis on data quality and post-training alignment. The release also came with an improved license and friendlier commercial terms. These details made it not just "another open model" but a product that combined open weights, frontier performance, and commercial usability all at once.

The community response was extraordinary. Researchers fine-tuned it as a base, startups deployed it privately, and individual developers ran the 8B version locally. Llama 3 became the de-facto standard of open models, with countless adaptation, quantization, and evaluation projects sprouting around it. A more concrete shift followed: enterprises no longer had to rely solely on closed APIs—open models became a genuinely viable alternative.

In July Meta released the 405B Llama 3 flagship, complete with an ecosystem of tooling. The 405B matched or exceeded contemporary closed rivals on many capabilities, turning "open models reach the frontier" from an accident into an established fact. Compared to Llama 2 in 2023, Llama 3 completed a qualitative leap—it was no longer the thing being "chased" but a participant that set the benchmark.

Llama 3's deep influence was making "open versus closed" a question the whole industry had to answer seriously. Closed vendors began publicly responding to the open challenge, and governments and academia entered a lasting policy debate about open weights. For ordinary developers, the most direct change was more choice: models were no longer the patent of a few companies but something you could download, modify, and deploy on your own servers.

Looking back at April 2024, Llama 3's significance is not a single score but the rewriting of the competitive baseline. It proved open models can reach the frontier, and that fact changed the direction of the whole AI ecosystem—the later rise of DeepSeek, Qwen, and other open models is continuous with the space Llama 3 opened. Open source was no longer "settling for less" but one of the mainstream routes running parallel to closed models.