Language modelLimited access
Koa
Salesforce
- Released
- -
- Data date
- October 7, 2026
- Input
- CRM tasks and synthetic workflow scenarios; input format not specified
- Output
- Step-by-step CRM actions and tool calls; output format not specified
Koa is Salesforce’s first reasoning model for CRM tasks. Salesforce post-trained NVIDIA Nemotron 3 Super on synthetic workflows from sales, service, and other business areas. Koa is designed to plan multi-step processes and select tools for actions.
Access is limited to selected Agentforce pilot customers. Salesforce expects general availability in U.S. regions in winter 2026.
Specifications and access
| Specification | Value and source |
|---|---|
| Model class | CRM reasoning model for multi-step business workflows and tool useSource |
| Input modalities | CRM tasks and synthetic workflow scenarios; input format not specifiedSource |
| Output modalities | Step-by-step CRM actions and tool calls; output format not specifiedSource |
| Base model | Post-trained from NVIDIA Nemotron 3 Super using synthetic CRM scenariosSource |
| Availability | Available only to selected Agentforce pilot customers; general availability in US regions is expected in winter 2026Source |
Page 1 of 3
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
Tau2Bench · Airline (Show measurement, test conditions, and source)
- Source value
- 62
- Score
- 62
- Metric
- pass@1, four-trial mean
- Unit
- %
- Benchmark version
- Salesforce Koa technical report Table 1
- Category
- agentic-and-enterprise
- Direction
- Higher is better
- Sample
- 50
- Source type
- vendor-reported
- Evaluator
- Salesforce Koa technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Salesforce Koa is the RL-trained checkpoint (the model states this explicitly). The paper evaluates all checkpoints in BF16. Tau2Bench uses GPT-4.1 as the user simulator and averages pass@1 over four trials; domain task counts are Airline 50, Retail 114 and Telecom 114.
Tau2Bench · Retail (Show measurement, test conditions, and source)
- Source value
- 81.6
- Score
- 81.6
- Metric
- pass@1, four-trial mean
- Unit
- %
- Benchmark version
- Salesforce Koa technical report Table 1
- Category
- agentic-and-enterprise
- Direction
- Higher is better
- Sample
- 114
- Source type
- vendor-reported
- Evaluator
- Salesforce Koa technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Salesforce Koa is the RL-trained checkpoint (the model states this explicitly). The paper evaluates all checkpoints in BF16. Tau2Bench uses GPT-4.1 as the user simulator and averages pass@1 over four trials; domain task counts are Airline 50, Retail 114 and Telecom 114.
Tau2Bench · Telecom (Show measurement, test conditions, and source)
- Source value
- 60.5
- Score
- 60.5
- Metric
- pass@1, four-trial mean
- Unit
- %
- Benchmark version
- Salesforce Koa technical report Table 1
- Category
- agentic-and-enterprise
- Direction
- Higher is better
- Sample
- 114
- Source type
- vendor-reported
- Evaluator
- Salesforce Koa technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Salesforce Koa is the RL-trained checkpoint (the model states this explicitly). The paper evaluates all checkpoints in BF16. Tau2Bench uses GPT-4.1 as the user simulator and averages pass@1 over four trials; domain task counts are Airline 50, Retail 114 and Telecom 114.
Tau2Bench · task-weighted average (Show measurement, test conditions, and source)
- Source value
- 69.41
- Score
- 69.41
- Metric
- task-weighted mean of pass@1, four-trial mean
- Unit
- %
- Benchmark version
- Salesforce Koa technical report Table 1
- Category
- agentic-and-enterprise
- Direction
- Higher is better
- Sample
- 278
- Source type
- vendor-reported
- Evaluator
- Salesforce Koa technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Salesforce Koa is the RL-trained checkpoint (the model states this explicitly). The paper evaluates all checkpoints in BF16. Tau2Bench uses GPT-4.1 as the user simulator and averages pass@1 over four trials; domain task counts are Airline 50, Retail 114 and Telecom 114.
BFCL · overall accuracy (Show measurement, test conditions, and source)
- Source value
- 66.63
- Score
- 66.63
- Metric
- accuracy
- Unit
- %
- Benchmark version
- Salesforce Koa technical report Table 1
- Category
- agentic-and-enterprise
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Salesforce Koa technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Salesforce Koa is the RL-trained checkpoint (the model states this explicitly). The paper evaluates all checkpoints in BF16. Tau2Bench uses GPT-4.1 as the user simulator and averages pass@1 over four trials; domain task counts are Airline 50, Retail 114 and Telecom 114.
BFCL · Non-Live AST · Average (Show measurement, test conditions, and source)
- Source value
- 84.85
- Score
- 84.85
- Metric
- accuracy
- Unit
- %
- Benchmark version
- BFCL Table 3, Salesforce Koa RL-only
- Category
- BFCL-subtask
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Salesforce Koa technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Appendix H Table 3, exact Salesforce Koa RL-only column for Non-Live AST / Average; all values are percentages and all checkpoints are evaluated in BF16. This is the RL-trained checkpoint; SFT-only and Nemotron base values are excluded.
BFCL · Non-Live AST · Simple (Show measurement, test conditions, and source)
- Source value
- 69.92
- Score
- 69.92
- Metric
- accuracy
- Unit
- %
- Benchmark version
- BFCL Table 3, Salesforce Koa RL-only
- Category
- BFCL-subtask
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Salesforce Koa technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Appendix H Table 3, exact Salesforce Koa RL-only column for Non-Live AST / Simple; all values are percentages and all checkpoints are evaluated in BF16. This is the RL-trained checkpoint; SFT-only and Nemotron base values are excluded.
BFCL · Non-Live AST · Multiple (Show measurement, test conditions, and source)
- Source value
- 94
- Score
- 94
- Metric
- accuracy
- Unit
- %
- Benchmark version
- BFCL Table 3, Salesforce Koa RL-only
- Category
- BFCL-subtask
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Salesforce Koa technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Appendix H Table 3, exact Salesforce Koa RL-only column for Non-Live AST / Multiple; all values are percentages and all checkpoints are evaluated in BF16. This is the RL-trained checkpoint; SFT-only and Nemotron base values are excluded.
BFCL · Non-Live AST · Parallel (Show measurement, test conditions, and source)
- Source value
- 87.5
- Score
- 87.5
- Metric
- accuracy
- Unit
- %
- Benchmark version
- BFCL Table 3, Salesforce Koa RL-only
- Category
- BFCL-subtask
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Salesforce Koa technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Appendix H Table 3, exact Salesforce Koa RL-only column for Non-Live AST / Parallel; all values are percentages and all checkpoints are evaluated in BF16. This is the RL-trained checkpoint; SFT-only and Nemotron base values are excluded.
BFCL · Non-Live AST · Parallel Multiple (Show measurement, test conditions, and source)
- Source value
- 88
- Score
- 88
- Metric
- accuracy
- Unit
- %
- Benchmark version
- BFCL Table 3, Salesforce Koa RL-only
- Category
- BFCL-subtask
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Salesforce Koa technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Appendix H Table 3, exact Salesforce Koa RL-only column for Non-Live AST / Parallel Multiple; all values are percentages and all checkpoints are evaluated in BF16. This is the RL-trained checkpoint; SFT-only and Nemotron base values are excluded.
BFCL · Live AST · Average (Show measurement, test conditions, and source)
- Source value
- 80.16
- Score
- 80.16
- Metric
- accuracy
- Unit
- %
- Benchmark version
- BFCL Table 3, Salesforce Koa RL-only
- Category
- BFCL-subtask
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Salesforce Koa technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Appendix H Table 3, exact Salesforce Koa RL-only column for Live AST / Average; all values are percentages and all checkpoints are evaluated in BF16. This is the RL-trained checkpoint; SFT-only and Nemotron base values are excluded.
BFCL · Live AST · Simple (Show measurement, test conditions, and source)
- Source value
- 86.43
- Score
- 86.43
- Metric
- accuracy
- Unit
- %
- Benchmark version
- BFCL Table 3, Salesforce Koa RL-only
- Category
- BFCL-subtask
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Salesforce Koa technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Appendix H Table 3, exact Salesforce Koa RL-only column for Live AST / Simple; all values are percentages and all checkpoints are evaluated in BF16. This is the RL-trained checkpoint; SFT-only and Nemotron base values are excluded.
Page 1 of 3
Sources and data date
| Type | Evidence and data date |
|---|---|
| Input | Documented modalities. Retrieved October 7, 2026. |
| Output | Documented modalities. Retrieved October 7, 2026. |
| Research status | Research date October 7, 2026. 1 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not published | demo-benchmark-scores Original wording The Salesforce Koa paper reports benchmark scores for the RL-trained Koa checkpoint. Product demonstrations without a controlled dataset, scoring method and numeric checkpoint result were not treated as additional benchmark evidence. |
| Additional source | Salesforce Koa technical report (retrieved October 7, 2026) |