NvidiaBrian Caulfield3 min readintermediate
‘Now We Can Know Everything and Do Anything,’ Jensen Huang Says at Dreamforce
Summary
Nvidia’s Jensen Huang announced Salesforce’s Koa, a CRM‑reasoning LLM built by fine‑tuning Nvidia Nemotron 3 Super on a synthetic, three‑decade‑spanning dataset. Koa uses supervised fine‑tuning plus RL (NeMo RL, Gym, AutoModel), covers 14+ industries, and claims 3× fewer errors on Salesforce’s CRM‑Bench versus leading models. It’s already in internal Slack agents and slated for limited customer p…
- Koa is a custom LLM for CRM tasks, built on open‑source Nemotron 3 Super via synthetic data, not real customer data.
- Training pipeline: supervised fine‑tuning → reinforcement learning with Nvidia’s NeMo stack (RL, Gym, AutoModel).
- Benchmark claim: 3× fewer errors on Salesforce’s CRM‑Bench (tasks like opportunity updates, case routing).
- Deployment roadmap: internal Slack agents now, customer pilots in Oct (e.g., Formula 1, UChicago Medicine), GA winter 2026 US.
Shows how large enterprises are internalizing LLM development with synthetic data to avoid data‑privacy issues, and how Nvidia’s open‑model ecosystem (Nemotron 3 Super + NeMo) is being leveraged for domain‑specific reasoning. The claimed error reduction hints at tangible productivity gains for CRM…
4/10
