🔶 Reflection AI 推出开源模型 Beam，瞄准成为中国实验室的西方替代方案

Reflection AI 是一家将自身定位为“美国对开源中国 AI 的回应”的初创公司，现正发布其首个模型 Beam。

这家由前 DeepMind 研究人员于 2024 年创立的公司，将 Beam 定位为一款“主力型”模型，其性能优于其他西方开源模型。联合创始人兼 CEO Misha Laskin 告诉 Semafor，该模型尤其擅长编程和 Agent 任务。他表示，与同类开源模型相比，Beam 在解决问题时所需的计算能力少 3 至 4 倍，因此运行速度更快、成本也更低。

Beam 的能力可与中国 Z.ai 近期发布的模型 GLM-5.2 相当，后者于 6 月推出；Beam 正在“接近”阿里巴巴最先进模型 Qwen 3.8-Max 的性能。Qwen 3.8-Max 仍落后于 OpenAI 和 Anthropic 最顶尖的闭源美国模型。

中国的初创公司和科技巨头在开源 AI 领域一直处于世界领先地位。多家西方大型公司近期开始采用中国模型，以降低客服、常规代码编写等特定重复性任务的成本。北京也将开源作为外交工具加以推广，帮助其他国家推进技术发展。但由于开源模型可以被下载并以较少的安全防护措施运行，它们的崛起也引发了安全担忧：美国和英国政府评估人员发现，近期的中国模型可能帮助黑客利用代码漏洞。

总部位于纽约的 Reflection 认为，对于希望构建主权 AI 系统、但不愿或无法使用中国模型，同时仍希望“拥有”并控制自身 AI 使用方式的企业、政府及其他实体而言，市场空间很大，Laskin 表示。

“他们如今实际上没有什么很好的选择，”他说。

Reflection 已开始训练下一代模型。Laskin 表示，该模型将比 Beam “强大得多”，性能也会显著提升。

Reflection AI, the startup billing itself as America’s answer to open-source Chinese AI, is releasing its first model, called Beam.

The company, founded in 2024 by former DeepMind researchers, is positioning Beam as a “workhorse” that performs better than other Western open models. The model is particularly strong at coding and agentic tasks, cofounder and CEO Misha Laskin told Semafor. It also needs three to four times less computing power to reason through a problem than comparable open models, he said, making it faster and cheaper to run.

The model’s capabilities are comparable to one of China’s Z.ai’s recent models, GLM-5.2, which was released in June; Beam is “approaching” the performance of Alibaba’s most advanced model, Qwen 3.8-Max, which trails the top closed American models from OpenAI and Anthropic.

China’s startups and tech giants have led the world in open-source AI. Several large Western companies have recently embraced Chinese models as a way to reduce costs for certain repetitive tasks like customer service and writing routine code. Beijing has also promoted open source as a diplomatic tool to help other countries in their tech development. But because open models can be downloaded and run with fewer safeguards, their rise has also raised security concerns: US and UK government evaluators found that recent Chinese models could help hackers exploit code vulnerabilities.

New York-based Reflection sees a large market in businesses, governments, and other entities building sovereign AI systems that don’t or can’t use Chinese models, but still want to “own” and control how they use AI, Laskin said.

“They don’t really have very good options today,” he said.

Reflection is already training its next model, which Laskin said would be “much more” powerful and performant than Beam.