Salesforce built Koa on open Nemotron weights, not a frontier API
Its CRM model matches frontier accuracy with 3x fewer errors — open weights are production-ready, and Nemotron 3 Super is yours to build on too.
Copy markdown
What Salesforce actually shipped
Koa is Salesforce's first in-house CRM reasoning model, post-trained on Nvidia's Nemotron 3 Super with supervised fine-tuning plus RL (GRPO) in NeMo. On Salesforce's own CRM benchmark it matches or beats leading models while making three times fewer errors on real actions like updating opportunities and routing cases. It's pilot-only today, with general availability expected in U.S. regions this winter.
The part you can grab today: Nemotron 3 Super
The base is open. Nemotron 3 Super is 120B total / 12B active — a hybrid Mamba-Transformer MoE with a native 1M-token context, on Hugging Face under Nvidia's commercial-use Open Model License. It posts LiveCodeBench 81.2, SWE-Bench 59.2 and AIME25 90.2, and ships inference cookbooks for vLLM, SGLang and TensorRT-LLM plus LoRA/RL fine-tuning templates. You can post-train your own domain agent the same way Salesforce did.
Why the frontier labs should notice
Salesforce's line is blunt: 'reasoning has always been something that we've relied on the frontier model providers for. Until now.' Building on open weights cut its token costs and kept customer data in-house — a template any team with a narrow, high-volume task can copy, and a direct hit to the enterprise-API revenue the big labs lean on.
Elsewhere: Claude Code 2.1.273
The latest build adds gateway hint headers (x-claude-code-request-class, x-claude-code-agent-type) for teams routing agent traffic through an LLM gateway, lets you fork sessions started with --remote-control, and hardens MCP disconnect handling. Mostly maintenance, but the headers matter if you meter or bill per agent.