Llama 4 发布

2025 年 4 月 5 日，Meta 发布 Llama 4 Scout 与 Maverick 的开放权重。发布页把它们称作同一“牧群”的成员：都使用混合专家架构，都能原生接收文本与图像。社区可以下载这两款模型，却不能下载家族中最大的 Behemoth。它仍在预览或训练阶段，同时已经在幕后向较小模型提供蒸馏信号。

这让 Llama 4 的“开放”从一开始就带着层次。Scout 约有 1090 亿总参数、16 个专家，每个 token 激活约 170 亿参数；Maverick 约有 4000 亿总参数、128 个专家，激活量同样约 170 亿。路由器只让一部分专家参与每一步计算，于是“权重文件有多大”和“生成一个 token 要算多少”不再是同一个问题。开发者得到更大的容量选择，也要承担专家并行、显存和推理框架适配的新复杂度。

Scout 最响亮的规格是 1000 万 token 上下文。Meta 的材料说明，模型在较短上下文上训练与后训练，再通过长度扩展方法和专门数据把窗口推远，并展示大海捞针与长代码类测试。这些证据支持“模型能够在官方设置下处理极长输入”，却不能把三个问题合并：内容能否塞进窗口、某个片段能否被找到、整个任务能否保持稳定判断。十万页材料装得下，不代表其中的矛盾、时间关系和隐含前提都被可靠理解。

Maverick 更强调综合能力，并使用 Behemoth 产生的软目标进行蒸馏。Meta 披露的 Behemoth 规模约为两万亿总参数、2880 亿激活参数，但发布时没有同步给出可用权重。于是社区拿到的不只是两个模型，也拿到一种新的开放结构：学生可以部署和微调，教师、完整训练数据与训练过程仍留在公司内部。开放权重扩大了可研究范围，却没有让整个生产链透明。

发布页还列出与 GPT-4o、Gemini、DeepSeek 等模型的比较。这些结果来自 Meta 选定的版本、提示和评测管线，应当被当作产品定位，而不是跨场景的独立决赛。真正落到部署端的问题更琐碎：10M 上下文是否值得相应内存与时延，MoE 路由在所用框架中是否高效，图像输入是否满足具体任务，许可证是否覆盖预期用途。

Llama 4 的故事因此不只关于“开放模型也有多模态和 MoE”。它更具体地展示了前沿模型家族如何被分层交付：最庞大的教师负责在幕后传递能力，较小的学生负责进入社区。Scout 与 Maverick 确实走出了围栏；Behemoth 留下的空位提醒人们，能够下载一个模型，与能够看见它完整的成长过程，仍是两种不同程度的开放。

Meta released open weights for Llama 4 Scout and Maverick on 5 April 2025. The launch presented them as members of one “herd”: both used mixture-of-experts architectures, and both accepted text and images natively. The community could download those two models. It could not download the largest family member, Behemoth, which remained in preview or training while already supplying distillation signals behind the scenes.

Openness in Llama 4 was layered from the beginning. Scout had about 109 billion total parameters and sixteen experts, activating roughly 17 billion parameters per token. Maverick had about 400 billion total parameters and 128 experts, while activating about the same 17 billion. A router selected only part of the expert capacity for each step, so the size of the stored weights and the computation required to generate one token were no longer the same number. Developers gained access to greater capacity and inherited new complexity in expert parallelism, memory, and inference-framework support.

Scout’s loudest specification was a ten-million-token context window. Meta described training and post-training at shorter lengths, followed by techniques and specialized data intended to extend the model, and showed needle-in-a-haystack and long-code tests. This supported a bounded claim: under the publisher’s settings, the model could operate over extremely long inputs. It did not collapse three separate questions into one. Material may fit in the window; a passage may be retrievable; coherent judgment over the entire task may still fail. Fitting the equivalent of many thousands of pages does not guarantee reliable understanding of their contradictions, chronology, and unstated assumptions.

Maverick emphasized broader capability and used soft targets from Behemoth during distillation. Meta described Behemoth at roughly two trillion total parameters and 288 billion active, but did not release its usable weights alongside the students. The community therefore received not only two models but a new structure of openness: students available for deployment and fine-tuning, while the teacher, full training data, and production process remained inside the company. Open weights widened the surface for study without making the entire supply chain transparent.

The launch also compared the family with versions of GPT-4o, Gemini, DeepSeek, and other systems. Those were Meta’s selected models, prompts, and evaluation pipelines—useful for positioning, not an independent final across every setting. Deployment turned on less theatrical questions: whether ten million tokens justified the memory and latency, whether MoE routing ran efficiently in the chosen stack, whether image input met the actual task, and whether the license covered the intended use.

Llama 4 was therefore more specific than “open models now have multimodality and MoE.” It showed how a frontier model family could be delivered in layers. The largest teacher transferred capability backstage; smaller students entered the community. Scout and Maverick did cross the fence. The empty place where Behemoth might have stood reminded users that downloading a model and seeing the full process that produced it remain different degrees of openness.