Salesforce post-trained Nvidia's open 120B into an enterprise agent and beat GPT-4.1 on tool use

Salesforce released Koa, built by post-training Nemotron-3-Super-120B, Nvidia's open-weight hybrid Mamba-Transformer MoE with 120B total and 12B active parameters (TechCrunch, paper at arXiv 2609.15066). The training was GRPO reinforcement learning on public and synthetic data only. No customer data, which Salesforce says explicitly and which matters for their enterprise buyers.

Koa beats its own base most clearly on multi-turn tool use, and surpasses GPT-4.1 on public tool-use, agentic-reasoning and CRM benchmarks. It stays below frontier models generally. Nobody is claiming otherwise.

Read that shape carefully, because the shape is the news. A CRM vendor took someone else's open base, spent RL compute instead of pretraining compute, and got a domain agent that beats a model from the lab that defined the category. No $500M training run. No cluster of their own.

Shanghai AI Laboratory did the same move on a different base. InternLM published Atria Dawn Preview to Hugging Face with MIT weights and a 1M-token context, repo live September 11, FP8 checkpoint September 12, paper September 14 with 143 signed authors (Hugging Face, arXiv 2609.15818). It takes Z.ai's 744B GLM-5.2 and post-trains it for research loops: problem analysis, tool use, code implementation, experiment execution, failure recovery. Competitive with frontier agents across 16 benchmarks, top reported score on five.

Two labs. Two different open bases. Same play, four days apart.

The Atria Dawn paper carries a second result I haven't stopped thinking about. The authors studied their own development process: 769 task records from 56 participants, and about a third of completed AI-assisted tasks were rated infeasible without AI. Agents proposed methods and implemented revisions while humans kept final decisions. That's a lab publishing the labor data for building the model alongside the model.

For anyone with a domain and no pretraining budget, the recipe is now written down twice in one week. Take an open base with a published architecture. Build a verifiable reward for your domain. Run GRPO. You need eval infrastructure and RL compute, not a datacenter.

The counterweight is Nathan Lambert's estimate that the open-closed gap runs 4 to 6 months. You're not building a frontier model. You're building something that beats last year's frontier model at one job, which for most enterprise work is the job.