Why We Post-Trained Our Own Reasoning Model
Salesforce, Wednesday, September 16th, 2026
Salesforce explains its decision to post-train a proprietary reasoning model for enterprise agent workloads.
Salesforce describes why it invested in post-training its own reasoning model instead of relying solely on general-purpose frontier models.
The post argues that enterprise agent tasks have distinctive requirements around tool use, structured output and predictable behavior that benefit from targeted training.
It outlines the data and evaluation approach used, and where the resulting model is applied inside the Agentforce stack. The piece also addresses cost and latency considerations in high-volume agent deployments.