欧美AI初创企业在开源模型领域掀起新一轮竞速。2026年10月3日至6日短短四天内，德国Aleph Alpha、美国Reflection AI、法国Mistral三家公司先后发布了三款开源模型1。其中，Aleph Alpha推出的Kolibri模型拥有780亿总参数1，Reflection AI发布的Beam模型规模达5010亿参数且每token激活230亿参数1，Mistral随后推出的Large 4模型则达到1.05万亿参数规模，每token激活520亿参数1。这波密集发布的模型均将主要对标矛头指向中国主流开源模型1。
Reflection AI明确将Beam定位为"面向中国开源模型的美国本土替代方案"1。这一开源浪潮与欧洲"主权AI"复兴相伴随，西方初创公司借助开源模式和监管合规优势争夺市场份额1。Mistral更是计划在欧盟境内建成1吉瓦的主权算力1。然而，这些企业仍面临用户迁移成本高和难以完全割离对中国技术依赖等挑战1。值得注意的是，斯坦福2026年AI指数报告显示，最好的闭源模型仅领先最好的开源模型3.3%1，这意味着开源技术与闭源方案的差距正在快速缩小。
Within four days in early October, three European and American artificial intelligence companies unveiled competing open-source models in rapid succession. On October 3, German firm Aleph Alpha released Kolibri, featuring 78 billion parameters 1. Two days later, on October 5, U.S.-based Reflection AI introduced Beam, a model with 501 billion parameters and 23 billion activated parameters per token 1. On October 6, French company Mistral followed with Large 4, boasting 10.5 trillion parameters and 52 billion activated parameters per token 1.
These releases represent a strategic shift in the open-source AI landscape, with each model explicitly designed as a Western alternative to dominant Chinese open-source offerings 1. Reflection AI positioned Beam specifically as "an American domestic alternative to Chinese open-source models" 1. This wave of model launches coincides with Europe's revival of "sovereign AI" initiatives, as startups leverage the open-source approach and regulatory compliance advantages to compete for market share 1. Mistral is planning to build one gigawatt of sovereign computing capacity within the European Union as part of this broader effort 1.
The competitive environment has shifted significantly in favor of open-source solutions. According to Stanford's 2026 AI Index, the best closed-source models hold only a 3.3 percent performance advantage over the best open-source models 1, narrowing the technological gap that previously favored proprietary systems 1.
评论
还没有评论，欢迎留下第一条。