Google 开源 Gemma

2024 年 2 月，Google DeepMind 发布了 Gemma 2B 与 7B 两个开源模型。单看参数规模，这个发布在 2024 年的大模型浪潮里不算耀眼——但它的象征意义很大：这是谷歌第一次向开源社区开放自家大模型的权重。那个长期被诟病「只出不进」的巨头，终于下场了。

在此之前的开源格局，是 Meta 的 LLaMA 系列和欧洲的 Mistral 在唱主角，谷歌一直以闭源的 Gemini 应战。开源社区对谷歌的态度相当复杂：研究论文、技术工具慷慨开放，模型权重却始终锁在自家产品里。Gemma 的发布，等于公开承认了开源模型已经是巨头必须布局的生态位，而不是可以无视的支流。

Gemma 的产品定位也很有讲究。2B 与 7B 的规模刻意面向个人开发者与端侧场景，可以在消费级硬件上运行，配套了模型卡、工具链与对开发者友好的许可。它不是要和 Gemini 抢旗舰的位置，而是去补谷歌缺失的另一条腿——让开发者生态里出现谷歌系模型的身影，让社区研究与第三方应用有谷歌的选择。

这个转向的意义远超 Gemma 本身。它标志着开源与闭源的关系，从「挑战者 vs 守成者」变成了巨头内部的双线作战：开源养生态、圈开发者，闭源做旗舰、担商业化。此后谷歌持续迭代 Gemma 系列，其他巨头也各自布局，开源模型正式成为大厂战略里的一根支柱。

回看 Gemma 的发布，它记录的不是某个模型的技术突破，而是整个行业竞争格局的一次显性化：当连谷歌都开始开源，开源模型就不再是任何公司可以忽视的力量，而是 AI 版图里与闭源旗舰并行的、必须认真经营的一条线。

In February 2024 Google DeepMind released Gemma in two sizes, 2B and 7B. By the parameter standards of 2024's model wave, the release was unremarkable—but its symbolism was large: it was the first time Google opened the weights of its own large models to the open community. The giant long accused of "give nothing back" had finally stepped in.

Before this, the open landscape was led by Meta's LLaMA line and Europe's Mistral, while Google answered with closed Gemini. The community's attitude toward Google had long been ambivalent: generous with papers and tools, but with model weights locked behind its own products. Gemma's release was a public admission that open models had become an ecosystem position a giant had to occupy, not a tributary to ignore.

Gemma's positioning was deliberate. The 2B and 7B sizes were aimed at individual developers and on-device scenarios, runnable on consumer hardware, with model cards, tooling, and developer-friendly licensing. It was not meant to compete with Gemini for the flagship crown, but to build the missing second leg—putting Google-family models into the developer ecosystem, letting community research and third-party apps have a Google option.

This pivot's significance goes far beyond Gemma. It marked the relationship between open and closed from "challenger versus incumbent" to a dual-track strategy inside giants: open models feed the ecosystem and court developers, closed models carry the flagship and the business. Google kept iterating Gemma, other giants laid out their own tracks, and open models formally became one pillar of Big Tech strategy.

Looking back at Gemma's release, it records not a single model's technical breakthrough but a moment the industry's competitive map became explicit: when even Google starts open-sourcing, open models are no longer a force any company can ignore—they are a line running in parallel with closed flagships, a lane that has to be operated seriously.