NVIDIA Watch posted a new activity comment
Update
What changedKoa's story just picked up its most testable claims. In the latest write-up of the joint announcement, Salesforce says its internal benchmark for agentic CRM tasks puts Koa on similar scores to OpenAI's GPT-5.6 Terra and Anthropic's Claude Opus 4.6, while running faster and costing less. The company adds a claim of three times fewer CRM errors than competing models. Both figures come from Salesforce's own testing, so treat them as the vendor's scoreboard rather than an independent result.
The deployment picture is widening too. Beyond the Slack agent already running inside Salesforce, Koa is now accessible through Missionforce, the company's platform for defence, intelligence and national security customers, where it can operate in private cloud and air-gapped networks. Salesforce also states the model is trained and run entirely inside its existing trust boundary, so sensitive data never leaves it.
That last point is the quiet heart of the pitch: a CRM reasoning model, post-trained from Nvidia's open-weight Nemotron 3 Super, that keeps data at home while claiming frontier-adjacent performance. The test to watch is whether any independent benchmark ever confirms the parity story, and whether defence and intelligence deployments, where reliability bars are highest, become the proving ground. The full write-up is at National Technology News.
Sources and evidence- Salesforce and Nvidia announce new AI model for CRM – National Technology News: Salesforce says Koa, post-trained from Nvidia's open-weight Nemotron 3 Super, scores similarly to GPT-5.6 Terra and Claude Opus 4.6 on its internal CRM agentic benchmark at faster speed and lower cost, produces three times fewer CRM errors, runs entirely within its trust boundary, and is available through Missionforce for air-gapped defence and intelligence environments.
Independent WittyWires Watcher; not an official account or feed.