At the recent Dreamforce technology conference, Salesforce officially launched its first reasoning large model, Koa. The model is built on the open-source weights of NVIDIA's Nemotron foundation and was jointly fine-tuned by the two companies, aiming to deeply empower core business scenarios such as sales, marketing, and customer service.

For a long time, enterprise-level intelligent transformation has often been accompanied by high costs. Major closed-source AI laboratories tend to have enterprises upload all internal documents, code, prompts, and user feedback to their closed-source models and agents, which not only increases usage costs but also brings potential data security and compliance concerns. The introduction of Koa provides enterprises with a completely different open-source weight technology path. It not only focuses on solving practical business tasks but also did not ingest any real customer data during the entire training process, completely eliminating the risk of leaking customer sensitive information.

image.png

In terms of architecture design and cost control, Koa demonstrates significant economic advantages through NVIDIA's Nemotron foundation. Compared to calling third-party closed-source large models, Koa consumes significantly fewer tokens when performing the same business tasks. In addition, by leveraging the AI gateway's automatic scheduling and routing mechanism, enterprises can flexibly distribute requests according to actual business needs. Furthermore, its security mechanisms are deeply embedded within the Salesforce platform, fully complying with the strict data management requirements of enterprise customers.

Previously, if an agent needed to handle long-process or multi-step reasoning tasks, the Agentforce platform usually had to forward the request through an AI gateway to external cutting-edge large models such as Claude or ChatGPT. Jayesh Govindarajan, Senior Vice President of Salesforce's AI business, stated that although the team had long planned to develop a high-performance enterprise reasoning model, it was difficult to realize due to the lack of pre-trained base models in the U.S. that could meet the first-tier performance and clear data traceability. The emergence of the Nemotron base broke this deadlock, as its training process used completely self-generated simulated business synthetic datasets, successfully simulating complex scenarios such as customer complaints and sales promotion.

Although Salesforce has launched its own Koa large model, it has not changed its strategy of diversified collaboration. The company recently announced a project called ClaudeForce with Anthropic, allowing enterprise users to directly use Claude as an AI interaction front-end while ensuring that all core business data remains securely stored within Salesforce's business record system.