Salesforce 与 Nvidia 的新推理模型，正是 AI 实验室应当警惕的一切

本周，在 Salesforce 举办的大型 Dreamforce 科技大会上，一款名为 Koa 的新 AI 模型成为该公司的重要发布之一。Koa 是 Salesforce 的首个推理模型，基于 Nvidia 的开放权重 Nemotron模型构建。两家公司共同对 Koa 进行了后训练，使其擅长销售、营销和客户支持相关任务。

Koa 是一个鲜明例子，说明企业界对 AI 的需求正在与前沿实验室所提供的产品分道扬镳。专有 AI 实验室更希望企业将文件、代码、提示词和反馈直接上传到其模型及 Agent 中，并为此投入数百万美元。

但借助这款模型，Salesforce 向企业客户提供了：

• 闭源前沿模型之外的开放权重替代方案
• 经过训练、可完成特定工作任务的模型，而不是用于解决不可能完成的数学题
• 未吸收任何真实客户数据、因而不会将这些数据泄露给其他人的模型
• 有助于降低 AI 支出的模型，因为它完成相同工作所需的 Token 更少
• 可根据需求通过 AI “gateway”自动路由的模型
• 以及能够遵循嵌入 Salesforce 系统中的客户数据要求和安全规范的模型。

Koa 将作为 Salesforce 在 Agentforce 平台中提供的其他模型的替代选项。客户可在该平台中构建 Agent，以处理回答客户服务问题、预约安排等重复性任务。

Salesforce AI 执行副总裁 Jayesh Govindarajan 对 TechCrunch 表示：“我们已经构建了许多小型、任务专用的语言模型，它们是 Agentforce 产品组合的一部分。但在推理方面，我们一直依赖前沿模型提供商。直到现在。”

在 Koa 出现之前，如果某个 Agent 需要对长期运行或多步骤任务进行推理，这些提示词会通过 Agentforce 的 AI gateway——即决定由哪个模型处理哪项请求的系统——被路由至 Claude 或 ChatGPT 等前沿模型。

Govindarajan 表示：“我们此前没有这样做、没有训练自己的企业级前沿模型，原因之一是：我们一直都想做，但挑战始终在于缺少一个可作为起点的预训练基础模型。在 Nemotron 出现之前，没有一个可用的、由美国自主掌控的预训练模型：第一，它必须可获得；第二，它必须处于最先进水平；第三，它必须具备清晰的数据来源。我们根本不知道 Qwen 是用什么数据训练的。”他所说的 Qwen，是由阿里巴巴推出的一款热门中国开放权重模型。

对这类模型进行后训练，意味着将其从通用系统转变为熟悉销售和客户支持知识的系统。为此，Salesforce 和 Nvidia 没有使用任何 Salesforce 客户的真实数据，而是构建了模拟客户行为模式的合成数据。

Govindarajan 解释说：“我们实际上模拟了一个客户服务环境，其中包括具备特定角色设定的客服专业人员、打电话进入客服中心的愤怒客户，以及试图促成交易的销售专业人员等各种场景。”

Koa 的目标是：相较于将相同任务发送给 Claude 或 ChatGPT，它能够更好地完成 Salesforce 客户希望 Agent 执行的工作任务，同时减少 Token 消耗、降低成本。

Nvidia 企业生成式 AI 软件副总裁 Kari Ann Briski 对 TechCrunch 表示，借助 Nemotron，“我们拥有一种独特的推理架构，能够实现 Token 高效率。这大致是你所需要的三位一体：自主 AI、首 Token 响应时间，以及高效推理——这一切都关系到 Token 经济性。”

不过，Salesforce 并未完全放弃 Anthropic 或 OpenAI。该公司刚刚宣布与 Anthropic 达成名为 ClaudeForce 的合作关系，使企业能够将 Claude 用作 AI 交互界面，同时其数据仍保留在 Salesforce 的记录系统中，并由 Salesforce 的基础设施提供安全保障。

A new AI model called Koa is one of the biggest announcements from Salesforce this week at its giant Dreamforce tech conference. Koa is the company’s first reasoning model, built on Nvidia’s open-weight Nemotron model. The two companies worked together to post-train Koa to excel at sales, marketing, and customer-support-related tasks.

Koa is a shining example of how the enterprise world’s needs for AI are diverging from what the frontier labs are offering. Proprietary AI labs would rather have enterprises uploading files, code, prompts, and feedback directly into their models and agents, and spending millions to do so.

But with the model, Salesforce is offering its enterprise customers:

• an open-weight alternative to closed frontier models
• a model trained to do specific work tasks (rather than to solve impossible math problems)
• one that has not ingested any actual customer data and therefore cannot leak it to others
• a model that helps reduce AI spending, since it uses fewer tokens to do the same work
• one that can be automatically routed through an AI “gateway,” depending on the need
• and a model that follows all of a customer’s data requirements and security embedded within Salesforce.

Koa will be provided as an alternative to the other models Salesforce offers in its Agentforce platform, where its customers build agents to handle rote tasks like answering customer service questions or scheduling appointments.

“We’ve built many small task-specific language models, which are part of Agentforce’s portfolio,” Jayesh Govindarajan, EVP of Salesforce AI, told TechCrunch. “But reasoning has always been something that we’ve relied on the frontier model providers for. Until now.”

Before Koa, if an agent needed to reason through a long-running or multi-step task, those prompts would be routed to a frontier model like Claude or ChatGPT through Agentforce’s AI gateway (the system that decides which model handles which request).

“One of the reasons we hadn’t done this before, train our own enterprise-grade frontier model — we always wanted to — but the challenge has always been the lack of a pre-trained base model to start with. Until Nemotron came along, there was no sovereign American pre-trained model that was available, one, and two, that was state of the art, and, three, that had clear data provenance. We have no idea what Qwen trains on,” Govindarajan said, referring to the popular Chinese open-weight model produced by Alibaba.

Post-training a model like this means taking it from a general-purpose system to one well-versed in sales and customer support knowledge, and to do that, Salesforce and Nvidia did not use any actual data from Salesforce’s customers. Instead, they crafted synthetic data that mimicked customers’ patterns.

“We actually simulated a customer service environment with a persona customer service professional, including irate customers that call into the customer service center, all the way to a sales professional who’s trying to close a deal,” Govindarajan described.

Koa is meant to be better at the work tasks Salesforce customers want an agent to do — and cheaper, in terms of tokens burned — than sending those same tasks to Claude or ChatGPT.

With Nemotron, “we have a unique architecture for inference to be token efficient,” Kari Ann Briski, Nvidia’s VP of Generative AI Software for Enterprise, told TechCrunch. “It’s kind of the trifecta of things that you need to have: sovereign AI, time to first token, efficient reasoning, for the tokenomics of it all.”

However, Salesforce isn’t exactly abandoning Anthropic or OpenAI. It just announced a partnership with Anthropic called ClaudeForce that allows companies to use Claude as their AI interface, while their data remains in Salesforce’s system of records, secured by its infrastructure.