Salesforce built its own reasoning model on NVIDIA's open weights

Salesforce opened Dreamforce in San Francisco on Tuesday with a model of its own. Koa is an enterprise language model post-trained from NVIDIA's open-weight Nemotron-3-Super-120B, and it will be offered inside the Agentforce platform as an alternative to the frontier models Salesforce currently rents from other providers. A 16-page technical report went up alongside it, and it is the more interesting of the two documents — because it publishes a benchmark table that does not flatter the launch.

The report's own summary is the plainest description available: Koa "improves over its open-weight base, with the clearest gains on multi-turn tool use, and surpasses a strong proprietary baseline while remaining below the strongest frontier models." That is a narrower claim than the framing around the announcement, and it is worth taking at face value.

What Koa is

Koa starts from NVIDIA's Nemotron-3-Super-120B, an open-weight checkpoint NVIDIA published in March under its Nemotron Open Model License. The model card describes a hybrid latent mixture-of-experts architecture — interleaved Mamba-2 and MoE layers with some attention — carrying 120 billion total parameters with 12 billion active per token, and a context window of up to one million tokens. The low active-parameter count is the part that matters commercially: it is what makes a self-hosted 120B model plausible against per-token API calls to a frontier lab.

Salesforce post-trains that base with Group Relative Policy Optimization, the reinforcement-learning method popularized by DeepSeek-R1. The report states that training used only public and synthetically generated data, with no customer data — a claim the company will be asked to stand behind, and one that is easier to make when the training tasks are simulated rather than harvested. Training ran on a Slurm cluster of five NVIDIA B200 nodes, with the reward design first worked out on a cheaper ~30B Nemotron Nano model before being ported to the 120B policy.

The benchmark table, in full

Salesforce evaluated Koa against its own base and three proprietary models on two public tool-calling benchmarks and one enterprise benchmark of its own. The numbers, as published:

• Tau2Bench (multi-turn customer service, task-weighted average): Koa 69.41, Nemotron-3-Super-120B 68.64, GPT-4.1 54.48, Claude Opus 4.8 74.00, GPT-5.5 83.99.
• BFCL (agentic tool use): Koa 66.63%, base 64.73%, GPT-4.1 53.96%, GPT-5.5 67.63%, Claude Opus 4.8 78.18%.
• CRM Bench (Salesforce and Agentforce workflows, weighted average): Koa 0.86, base 0.84, GPT-4.1 0.81, Claude Opus 4.8 0.87, GPT-5.5 0.90.
• CRM Bench function-call accuracy, the clearest single gain: Koa 0.77 against the base model's 0.71.

Read that table without the press release and three things stand out. The margin over the open-weight base is under a point on Tau2Bench and about two points on BFCL — real, repeatable, and modest. The margin over GPT-4.1 is large, but GPT-4.1 is a 2025 model and not what anyone is deploying agents on today. And on every one of the three benchmarks, including the one Salesforce built and scored itself, a current frontier model is still ahead.

The method is the part worth copying

What distinguishes the work is not the score but where the training data comes from. Agentforce agents are configured in Agent Script, Salesforce's declarative language for describing an agent's router, subagents, typed actions, tool scopes and workflow instructions. Salesforce compiles those same specifications into a typed workflow graph — argument schemas, declared state effects, routing conditions, termination rules — and instantiates it as an executable simulated environment with state that tools read and mutate.

A simulation pipeline then expands each specification into scenario- and persona-conditioned multi-turn sessions, executed as online rollouts in NVIDIA's NeMo Gym, with rewards grounded in whether the model actually resolved the task through correct tool use rather than whether its text looked right. For public tool-use domains, where no Agent Script spec exists, Salesforce synthesizes an equivalent workflow graph so the same machinery drives both.

The consequence is a direct loop between agent authoring and model post-training: the artifact a customer writes to configure an agent is the same artifact that generates the training tasks and defines the reward. Every vendor with a declarative agent format — and there are several now — has the same loop available to it.

Why an applications company wants a reasoning model

We've built many small task-specific language models, which are part of Agentforce's portfolio. But reasoning has always been something that we've relied on the frontier model providers for. Until now.

That is the structural point. An applications vendor that routes every agent turn to an external lab pays per token, inherits that lab's deployment constraints, and has to explain the data path to regulated customers. Owning a specialized model changes the arithmetic on all three, even if the model is weaker in the abstract — because the tasks it has to do are not abstract. NVIDIA's Kari Ann Briski, VP of generative AI software for enterprise, framed the appeal to TechCrunch as "sovereign AI, time to first token, efficient reasoning, for the tokenomics of it all."

What is not stated

Salesforce has not said whether Koa's weights will be released, what it will cost inside Agentforce, or when it becomes generally available; TechCrunch reports only that it will be offered as an alternative to the platform's other models. The token-efficiency claim — that Koa burns fewer tokens for the same work than routing to a frontier model — is the company's, and no measurement supporting it appears in the technical report.

The report also flags its own main limitation: because Nemotron-3-Super-120B is itself already reinforcement-learning post-trained, Salesforce's finding that RL beats supervised fine-tuning for multi-turn tool use may not hold starting from a pre-RL checkpoint. The authors say so explicitly rather than generalizing. And CRM Bench, the benchmark on which Koa comes closest to frontier performance, is Salesforce's own — home-field evaluation, scored at home.

None of that makes the release less significant. The claim worth watching is not that Koa beats anyone. It is that an open-weight base with 12 billion active parameters, specialized with a company's own agent specifications, lands close enough on the narrow set of tasks that company actually sells. If that reproduces outside Salesforce's own evaluation, the interesting number is not on the benchmark table at all — it is the size of the bill that stops going to a frontier lab.