Meta 发布 Llama 3.1 405B

2024 年上半年，一条不成文的规矩横在开源与闭源之间：开源模型可以做得不错，但真正的旗舰能力，只存在于 GPT-4o、Claude 这些闭源 API 里。Meta 在 4 月发布了 Llama 3 的 8B 和 70B，很好用，但依然属于「中端」。Meta 想动那条规矩。

2024 年 7 月 23 日，Llama 3.1 发布。这一次 Meta 拿出的是一张前所未有的牌：405B。这是当时最大的开源模型，在多项基准上逼近甚至追平 GPT-4o。更重要的是，Meta 不是只发一个模型——8B、70B、405B 三档齐发，配套 Llama Stack 工具链和 128K 上下文的训练配方，还把许可改成了允许蒸馏。

同一天，扎克伯格发表了那篇著名的长文《开源 AI 是通往未来的正确道路》。他给出的理由很现实：开源让生态更繁荣、让 Meta 避免被单一供应商卡脖子、让技术民主化。批评者说这是商业算计，但无论如何，Meta 用行动把「开源也能做旗舰」变成了现实。

405B 的开源立刻点燃了整个生态。全球团队第一次能在旗舰级模型上做微调、做私有化部署、做二次创新。OpenAI 们的闭源护城河，第一次被正面凿开了一个大口子。此后「开源与闭源差距缩小」成为行业反复讨论的话题，源头都可以追溯到这一次发布。

回看 Llama 3.1，它既是技术事件，也是意识形态事件。它让「最强开源模型」与「最强闭源模型」在同一个赛场上短兵相接，也让开源从「够用就行」走向「正面对抗」。当后来 DeepSeek、Qwen 等开源模型不断刷新纪录时，它们其实都站在 2024 年 7 月这个转折点上——那一次，Meta 替整个开源世界，向闭源霸主们宣了战。

In the first half of 2024, an unwritten rule stood between open and closed source: open models could be decent, but true flagship capability lived only inside closed APIs like GPT-4o and Claude. Meta released Llama 3's 8B and 70B in April—good models, but still "mid-range." Meta wanted to move that rule.

On July 23, 2024, Llama 3.1 arrived. This time Meta played a card nobody had seen: 405B. It was the largest open model of its time, approaching or matching GPT-4o on many benchmarks. More importantly, Meta did not just release one model—8B, 70B, and 405B launched together, paired with the Llama Stack tooling, a 128K-context training recipe, and a license revised to permit distillation.

The same day, Zuckerberg published the famous essay "Open Source AI Is the Path Forward." His reasons were practical: openness makes ecosystems thrive, keeps Meta from being squeezed by a single supplier, and democratizes technology. Critics called it business calculation, but either way, Meta had turned "open source can also do flagship" into reality.

The open-sourcing of 405B immediately ignited the ecosystem. Teams worldwide could, for the first time, fine-tune, privately deploy, and innovate on a flagship-grade model. A gaping hole had been cut, for the first time, in the moat of the closed players. The recurring industry debate about the narrowing open/closed gap can all be traced to this release.

Looking back, Llama 3.1 is both a technology event and an ideological one. It put the strongest open model and the strongest closed model on the same battlefield, and moved open source from "good enough" to "frontal confrontation." When later open models like DeepSeek and Qwen kept setting records, they were all standing on this turning point of July 2024—the moment Meta declared war, on behalf of the entire open world, against the closed-source powers.