Koa is an enterprise language model announced by Salesforce and NVIDIA on September 15, 2026 at Salesforce's Dreamforce conference, described by Salesforce as its first CRM reasoning model for its Agentforce agent platform. It was built by post-training NVIDIA's open-weight Nemotron 3 Super foundation model on synthetic data modelled on Salesforce's CRM deployments, and is aimed at multi-step tool use inside customer-relationship workflows rather than at general-purpose chat (Source: salesforce.com). A technical report describing the training method was posted to arXiv on September 14, 2026 (Salesforce Koa: An Enterprise Language Model for Agentic Tool Use (technical report)). The report states that Koa improves on its base model, most clearly in multi-turn tool use, and exceeds GPT-4.1 while "remaining below the strongest frontier models" on the benchmarks it reports (Salesforce Koa: An Enterprise Language Model for Agentic Tool Use (technical report)).
Infobox
| Field | Value |
|---|---|
| Developer | Salesforce, in collaboration with NVIDIA |
| Announced | September 15, 2026 (Dreamforce) |
| Base model | NVIDIA Nemotron 3 Super (named "Nemotron-3-Super-120B" and "Nemotron-3-Super v3 (∼120B)" in the technical report) |
| Post-training | Reinforcement learning with Group Relative Policy Optimization (GRPO); the press release names supervised fine-tuning followed by GRPO, while the technical report states that RL was applied directly to the base model and Koa built on the RL-only checkpoint |
| Open weights | No — Salesforce states it controls the weights and runs inference within its own infrastructure |
| Availability | Select pilot customers in Agentforce at announcement; general availability stated for winter 2026 in U.S. regions |
| Technical report | arXiv:2609.15066, submitted September 14, 2026 (Salesforce Koa: An Enterprise Language Model for Agentic Tool Use (technical report)) |
Training and architecture
Koa is not trained from scratch. Salesforce post-trained NVIDIA's open-weight Nemotron 3 Super model. The press release states that Salesforce applied supervised fine-tuning and reinforcement learning with Group Relative Policy Optimization using NVIDIA's NeMo RL, NeMo Gym and NeMo AutoModel tooling (Source: salesforce.com). The technical report describes the same base model and optimization algorithm, and states that the model was "trained on public and synthetically generated data, with no customer data," but gives a different training sequence: RL was applied "directly to the base model rather than to an SFT checkpoint," supervised fine-tuning was a preliminary study, and "we build Salesforce Koa on the RL-trained checkpoint." In that study an SFT-only checkpoint trained on synthetic tool-use trajectories scored slightly higher than the RL checkpoint on Tau2Bench and Salesforce's CRM Bench but did not improve BFCL multi-turn accuracy, which the authors give as their reason for building on the RL checkpoint. They caution that the base model was itself already RL post-trained, so that "starting from a pre-RL checkpoint could alter the conclusion" (Salesforce Koa: An Enterprise Language Model for Agentic Tool Use (technical report)).
The report identifies the distinctive component as a simulation-to-reward pipeline that "expands workflow specifications into persona-conditioned multi-turn tasks with task-resolution rewards grounded in successful tool use for data-dependent requests." For enterprise domains those specifications are written in Agent Script, which Salesforce describes as its declarative language for building Agentforce agents; for public tool-use domains the workflow structure is synthesized directly. The report states that the same simulation and grounded-reward machinery drives GRPO across both (Salesforce Koa: An Enterprise Language Model for Agentic Tool Use (technical report)).
In the report's setup, a specification is compiled into an executable environment in which a frozen helper model, Nemotron-3-Nano of about 30 billion parameters, plays the simulated customer, emulates tools that lack a production backend, and judges whether each numbered sub-question of the task was resolved. The reward is the share of sub-questions resolved, set to zero after two identical consecutive tool calls, and a sub-question needing customer or system data counts as resolved only if backed by a successful tool call. Training environments covered Salesforce support, healthcare administration and real estate, plus a workplace-assistant sandbox scored by exact end state. A separate single-step mode trains turns in which the correct action is to tell the user of a limitation rather than call a tool; the authors write that the base policy "largely lacks this skill," at times "fabricating confirmations for actions that never occurred." The recipe was developed on the Nano model, on which the chosen reward peaked early and collapsed, before being run stably on the Super model; training used five NVIDIA B200 nodes (Salesforce Koa: An Enterprise Language Model for Agentic Tool Use (technical report)).
Salesforce's announcement describes the training corpus as built entirely from synthetic scenarios simulating enterprise workflows across more than 14 industries, including manufacturing, financial services, healthcare and travel, each scenario pairing a persona with tasks and then mapping the sequence of actions and tool calls required to complete them. The company states that no customer data was used in training and that no customer data crosses its trust boundary during inference, and attributes the model's specialization to "27 years of Salesforce CRM intelligence" being placed "inside the model itself" (Source: salesforce.com).
Capabilities and benchmarks
The technical report states that 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," and that the results show "specification-driven reinforcement learning is a practical path to specializing open-weight foundation models for enterprise agentic tasks" (Salesforce Koa: An Enterprise Language Model for Agentic Tool Use (technical report)). Salesforce's announcement states that on its own CRM benchmark — a suite of tasks such as updating an opportunity, routing a case or scheduling a follow-up — Koa "already matches or exceeds leading model performance on CRM actions with three times fewer errors" (Source: salesforce.com). The comparison set behind that claim is not named in the announcement.
The report compares Koa with its base model, Nemotron-3-Super-120B, and with OpenAI's GPT-4.1 and GPT-5.5 and Anthropic's Claude Opus 4.8, all run or reported by Salesforce. On Tau2Bench, with GPT-4.1 as the simulated user over four trials, Koa's task-weighted average across airline, retail and telecom is 69.41 against 68.64 for the base, 54.48 for GPT-4.1, 74.00 for Claude Opus 4.8 and 83.99 for GPT-5.5; its telecom score is the same as the base's. On the Berkeley Function Calling Leaderboard benchmark it scores 66.63% against 64.73% for the base, 53.96% for GPT-4.1, 67.63% for GPT-5.5 and 78.18% for Claude Opus 4.8, with the largest gain over the base in the multi-turn category (59.50 against 54.12). On CRM Bench, which the report describes as single-turn Salesforce and Agentforce workflows scored on topic, function-call and free-text accuracy, Koa averages 0.86 against 0.84 for the base, 0.81 for GPT-4.1, 0.87 for Claude Opus 4.8 and 0.90 for GPT-5.5; its function-call accuracy of 0.77 is below that of all three proprietary models (Salesforce Koa: An Enterprise Language Model for Agentic Tool Use (technical report)). Per-domain and per-category scores with their settings are recorded on τ-bench and Berkeley Function Calling Leaderboard (BFCL), and the CRM Bench scores on the report's source page.
The report does not state the reasoning or prompt settings used for the proprietary models or the BFCL version. The CRM benchmark page that Salesforce's announcement links to describes scoring on accuracy (factuality, instruction following, conciseness and completeness), cost, speed, trust and safety, and sustainability, rather than the topic, function-call and text accuracy the report uses for CRM Bench (Source: salesforceairesearch.com).
Availability and deployment
The September 15 announcement was of a pilot rather than a general release. Salesforce's availability statement reads "Available to select pilot customers now in Agentforce; general availability expected winter 2026 in U.S. regions." At announcement Koa was in use inside Salesforce, including an agent in Slack for employee tasks, and was moving into customer pilots with 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero (Source: salesforce.com). The technical report does not discuss availability (Salesforce Koa: An Enterprise Language Model for Agentic Tool Use (technical report)).
The same announcement extends the Salesforce–NVIDIA collaboration to Missionforce, Salesforce's government and regulated-industry line, bringing Nemotron-based models and NVIDIA accelerated computing into private clouds, classified networks and air-gapped systems. Post-trained NVIDIA models powering Missionforce Operations were stated as available to select customers in October 2026 (Source: salesforce.com).
Open questions
- Whether Koa is the model's original name. The announcement's social-preview image is filed under a path naming a reasoning model called "Cerebro," and no source states a rename.
- The model's parameter count as shipped. The technical report names the base model "Nemotron-3-Super-120B" and describes the policy as "Nemotron-3-Super v3 (∼120B)," while the description of Nemotron 3 Super at NVIDIA Nemotron gives approximately 100 billion total parameters with up to 10 billion active per token, the figure from NVIDIA's December 2025 announcement. Whether the report's figure reflects the released size of Super, a later revision, or a naming convention is not established from the sources held; neither source gives Koa's active-parameter count.
- Which models Salesforce's "leading model performance" comparison covers on its CRM benchmark, and whether the three-times-fewer-errors figure is drawn from the same evaluation as the technical report's CRM Bench, on which Koa's average (0.86) is below Claude Opus 4.8's and GPT-5.5's.
- Whether supervised fine-tuning was part of the shipped model's training. The press release names SFT and GRPO; the technical report states that Koa is the RL-only checkpoint.
- Whether the weights will be released. Coverage of the announcement refers to open-model plans; the press release states only that Salesforce controls the weights.
Relationships
- depends-on: NVIDIA Nemotron — Koa is a post-trained derivative of Nemotron 3 Super
- depends-on: Post-Training — built by reinforcement-learning post-training (GRPO) rather than pretraining; the press release also names supervised fine-tuning
- related: Salesforce Koa: An Enterprise Language Model for Agentic Tool Use (technical report) — the technical report on the model's training and evaluation
- related: τ-bench, Berkeley Function Calling Leaderboard (BFCL) — public benchmarks on which the report's results are recorded
- related: Synthetic Data / Model-Generated Training Data — training tasks and personas are model-generated from agent specifications
- related: Salesforce, Inc. — developer and distributor
- related: Nvidia & TSMC — AI Compute Infrastructure — base-model provider and co-announcer
- instance-of: Agentic AI — the model's stated target is multi-turn tool use by agents
- related: Open-Weight Frontier Models — an instance of a proprietary product built on an open-weight base
- deploys-in: Financial Services — AI Deployment — named pilot customers include Baxter Credit Union and 1-800Accountant
- deploys-in: Healthcare — AI Deployment — named pilot customers include UChicago Medicine