欧洲开源数据库服务提供商 Aiven 表示，通过转向开放权重模型，其内部 AI token 成本降低了 80%，并正在构建一个网关，帮助开发者根据需要灵活切换模型。这家总部位于芬兰的公司以托管和提供开源数据库即服务为业务基础，过去几年拓展了机器学习领域。Aiven 战略副总裁 Ville Lehto 在接受《The》采访时表示，在处理开源 AI 基础设施时，要打造能在不同云平台间运行的产品并非易事。
European open source database service provider Aiven says it has saved 80 percent of its AI token costs internally by moving to open-weight models and is building a gateway to help devs swap models in and out as required. The Finland-based company built its business on hosting and providing open-source databases as a service, and in the last couple of years has moved into machine learning. Speaking to The, Ville Lehto, Aiven’s VP of strategy, said that when dealing with open source AI infrastructure, it's not trivial to make something that can run across the different clouds.
然而，在给开发者提供大语言模型（LLM）访问权限时，公司希望自行管理 token 成本，并掌握“在 AI 使用方面的自主命运”。
However, in giving its developers access to LLMs, the company wanted to manage its own token costs and control its "own destiny when it comes to AI usage."
在一波关于所谓“token 最大化利用”的讨论之后，开发者普遍更有兴趣通过混合使用多种模型来降低成本，包括比西方前沿模型更便宜的中国替代模型，以及成本更低的开放权重模型。开放权重模型公开发布了学习到的参数，主要是其权重和偏置，可以在本地运行或以比 OpenAI 和 Anthropic 等前沿模型更低的成本托管。例子包括 Meta 的 Llama 系列、阿里云的 Qwen 系列、中国的 DeepSeek、Mistral AI 的 Medium 3.5 和 Devstral，以及 Google 的 Gemma。
After a wave of debate around so-called token-maxxing, developers have generally become more interested in reducing their costs by using a mix of models, including cheaper Chinese alternatives to the Western frontier models, and open-weight models which come at a lower cost. Open-weight models have publicly released learned parameters, principally their weights and biases, and can be run locally or hosted at lower cost than the frontier models from OpenAI and Anthropic. Examples include the Llama Family from Meta, Qwen Family from Alibaba Cloud, China’s DeepSeek, Mistral AI’s Medium 3.5 and Devstral and, Google's Gemma.
Lehto 表示：“在 token 使用方面，正如我们今年早些时候所见，极少数用户是绝对的重度用户，就他们的支出而言，通过实际使用开放权重模型可以节省很大一部分。”他说，Aiven 面临一个选择：要么制定严格的 token 预算，要么尝试寻找更高效的 token 使用方式。选择后者后，公司实施了一个内部网关，帮助开发者更轻松地使用开放权重模型。Lehto 说：“我们希望让这个过程尽可能无摩擦。”
"When it comes to this token use, as we witnessed early on this year, there are a fraction of users that are absolutely heavy users, and when it comes to their spend, a lot of it could be saved by actually using open-weight models," Lehto said. He said Aiven faced a choice to either have strict token budgets or try to find a more efficient way of using tokens. Opting for the latter, it implemented an internal gateway to help developers work with open-weight models more easily, Lehto said. "We wanted to make it as frictionless as possible.
我们可以为员工提供不同的开放权重模型，无论是西方模型还是来自亚洲的模型，并与前沿模型竞争。我们一直专注于重度用户，在他们身上我们看到支出大约下降了 80%。不过，由于开放权重模型和前沿模型的性能各异，使用其他指标（如每项任务的成本）也很重要，他说。Aiven 围绕名为 LiteLLM 的开源框架构建了其网关，并对其进行托管。
We can offer our employees different open-weight models, be it Western models or the ones from Asia, and take on the frontier models. We've been focusing on the heavy users, and with them we see roughly 80 percent decline in spend." However, since the performance of open-weight and frontier models varies, it was also important to use other metrics, such as the cost per task, he said. Aiven built its gateway around the open source framework called LiteLLM, which it hosted.
它增加了预算管理、访问管理等新功能，并添加了其他开源组件。它在租用的 GPU 上运行模型，也自行托管模型。Lehto 表示，Aiven 继续使用前沿模型，但仅用于帮助规划工作。“在实际做事时，我们大量使用开放权重模型。大部分工作实际上是用开放权重模型完成的，但工作中最关键的部分是用前沿模型规划的。这是我们目前看到的最成功的方式。”
It added new features such as budgeting, access management, and added other open-source components. It runs the models on rented GPUs and self-hosts the models as well. Lehto said Aiven continued to use frontier models but only to help plan work. "When it comes to actually doing things, we use a lot of open-weight models. The majority of the work is actually done with open-weight, but the most critical bits of the work are planned with frontier models. That's how we see most success at the moment."
Aiven 计划将该网关作为一项服务提供给客户，通过让客户更容易在工作流中替换开放权重模型，帮助他们管理令牌成本。“既然我们的传承在于开源，这对我们来说感觉非常自然……作为一项服务提供给人们，因为它也符合企业需求。”Aiven 的客户包括德国电商公司 Conrad、英国支付处理公司 Dojo、网络安全供应商 Sophos 和通信公司 Avaya。它将只是众多试图帮助组织降低使用 AI 成本的供应商之一。®
Aiven plans to offer the gateway as a service to customers to help them manage token costs by making it easier to swap open-weight models in and out of their workflows. "Since our heritage is in open source, it's something that we feel very natural… offering as a service for people in the sense that it also fits the enterprise needs." Aiven's customers include German ecommerce company Conrad, UK payment processing company Dojo, cyber security vendor Sophos, and communications firm Avaya. It is set to be just one of the vendors positioning themselves to try to help organizations reduce the costs when using AI. ®