Back to News
ProductAI Understanding briefing

Salesforce launches Koa, a vertical AI model developed with NVIDIA

Salesforce has introduced Koa, a vertical AI model built on NVIDIA's Nemotron, designed to handle specific CRM workflows while reducing reliance on third-party providers.

4 min readRead the linked source
Source-page capture accompanying Salesforce launches Koa, a vertical AI model developed with NVIDIA
Source referenceSource recorded
Publisher
eu.36kr.com
Source link
eu.36kr.comhttps://eu.36kr.com/en/p/3992455185972230
Source type
Linked source — primary-source status has not been established.
ContextUnderstand this in 60 seconds

Start here

Key terms

Large Language Model (LLM)
A language model trained on massive text corpora to generate and analyze text.
Memory (Agent Memory)
Stored context an AI agent uses across steps or sessions to improve continuity.
Synthetic Data
Artificially generated data used to augment, simulate, or protect sensitive training data.
Test yourselfAI Models Explained Quiz

What happened

At the Dreamforce conference, Salesforce announced the launch of Koa, a vertical large language model developed in collaboration with NVIDIA. Koa is post-trained using NVIDIA's open-weight Nemotron model and is specifically optimized for Salesforce CRM tasks, such as updating opportunities and scheduling follow-ups. Salesforce claims that Koa improves tool-calling accuracy by 11%, enhances context recall reliability by 2.1 times, and increases long-conversation memory by 15% compared to its previous internal benchmarks. The company explicitly stated that Koa was trained using rather than raw customer data to address privacy concerns.

Salesforce unveiled Koa at its Dreamforce conference, positioning it as a vertical model designed to handle specific CRM workflows. The model is post-trained on NVIDIA's Nemotron open-weight architecture.

According to Salesforce's internal CRM Bench evaluation, Koa achieved a weighted average score of 8.6, trailing OpenAI's GPT-5.5 (9.0) and Claude Opus 4.8 (8.7), but outperforming GPT-4.1 (8.1).

Salesforce emphasized that Koa's training corpus consists entirely of synthetic scenarios simulating CRM workflows, asserting that no raw customer data was used in the process. This approach is intended to address data privacy concerns that have surfaced in the broader AI industry.

The company also announced a separate, concurrent partnership with Anthropic called 'Claudeforce,' which allows users to integrate Claude directly into Salesforce workflows, suggesting a hybrid strategy that utilizes both proprietary vertical models and external cutting-edge AI.

Source details: eu.36kr.com

Why it matters

The launch of Koa represents a strategic shift for SaaS companies attempting to balance the use of general-purpose AI with the need for data sovereignty and cost control. By developing a vertical model, Salesforce aims to mitigate the risk of data leakage associated with third-party model providers and reduce long-term dependency on expensive, closed-source AI services. This move highlights a growing industry trend toward 'Enterprise Sovereign AI,' where companies leverage open-weight models and proprietary business data to create specialized, cost-effective AI infrastructure. The partnership also underscores NVIDIA's broader strategy to provide the engineering toolchains and models necessary for enterprises to build their own custom AI solutions, potentially altering the competitive landscape between SaaS providers and AI labs.

The development of Koa reflects a defensive and offensive strategy for SaaS providers: by building vertical models, they protect their role as indispensable infrastructure for clients who might otherwise bypass software companies to build their own AI agents.

By reducing reliance on external, high-gross-margin AI models, Salesforce aims to improve its own margins and pass benefits to customers, while simultaneously addressing the 'black box' risks associated with third-party model providers.

NVIDIA's involvement, including CEO Jensen Huang's endorsement, signals a push to enable software companies to become 'AI companies' by providing the necessary hardware and software stack, effectively expanding NVIDIA's influence beyond the limited number of major AI research labs.

Interactive Mechanism

Interactive Mechanism: How It Actually Works

Explore the underlying technology behind this development interactively.

Thinking Budget (Test-Time Tokens):1,024 tokens
Complex Accuracy79%Math & Code Logic
Latency3.2sTime to first full output
Inference Cost$0.0092Per query estimated
Reasoning StyleStep VerificationInternal chain depth
Active Thinking Trace:
1Deconstruct user problem into formal constraints
2Propose candidate hypotheses & step-by-step calculation
3Self-correction: Backtrack and refute subtle edge cases
4Exhaustive consistency check & final output synthesis
Core takeaway: Test-time compute fundamentally changes AI economics. Instead of only scaling during pre-training, giving reasoning models more tokens at inference time allows them to systematically solve PhD-level STEM problems.
Interactive Concept Check+10 Points
AI Models Explained Quiz

What is the best response when AI Models Explained makes a mistake in production?

What to watch next

Industry observers should monitor whether other SaaS companies can successfully replicate Salesforce's vertical model strategy, given the significant requirements for high-quality training data and specialized engineering talent. While Salesforce has the scale to absorb the costs of post-training, smaller software firms may face barriers to entry. Additionally, the long-term impact on the 'hybrid model' approach—where enterprises route standardized tasks to vertical models while reserving complex, cross-domain reasoning for top-tier closed-source models—remains to be seen. The evolution of the talent pool for post-training and the potential for further 'de-NVIDIA-izing' by major AI labs like Anthropic are also key factors in this shifting ecosystem.

The primary challenge for the broader adoption of vertical models is the scarcity of specialized talent capable of high-quality post-training, a resource currently concentrated in top AI labs.

The sustainability of this model depends on whether the market demand for vertical AI can drive down the costs of development and training, similar to the historical growth of the mobile app development ecosystem.

The industry will likely move toward a hybrid model where enterprises route standardized, data-sensitive tasks to vertical models like Koa, while continuing to utilize general-purpose models for complex, cross-domain reasoning tasks.

Related guides & quizzes

AI Models ExplainedAI AgentsAI TrainingFuture of AITest what you know — try a free AI quizLook up an AI term in our glossary
Found this useful?