Deel is building specialist AI models around the employment and payroll knowledge it uses across more than 150 countries. Its Pangaea family includes K1 for knowledge and P1 for payroll, alongside an evaluation lab called Emet. The aim is to make frequently used expertise available quickly inside the product, while using retrieval to check details and supply current sources.[1][2]
K1 begins with a 9-billion-parameter open-weight model. Deel says its knowledge base contains about 43,500 pages and 69 million tokens, covering the policies and procedures its employees consult. Its technical account describes a memory-first system: training supplies a working understanding of the domain, while a lighter search step handles specifics, recent changes and citations.[2]
Teaching the same fact in several forms
Simply feeding documents to a model was not enough. Deel says its initial training corpus gave too little varied exposure to information in longer passages. It rebuilt the material into multiple views of each chunk, including restatements for different readers, numerical descriptions, summaries and question-and-answer examples. The engineering question was how reliably the model could use the knowledge, rather than how much text it had seen.[2]
In Deel’s internal evaluation, answering from memory took under eight seconds at the median, compared with 13–28 seconds for eight models using search over the knowledge base. Those timings measure completion of the answer. The company separately says production answers begin appearing within one second.[2]
What Deel’s figures measure
| Measure | Reported result | Meaning |
|---|---|---|
| Knowledge corpus | 43,500 pages; roughly 69 million tokens | Internal material used to teach domain knowledge |
| Full answer time | Under 8 seconds from memory | Median in Deel’s evaluation, not time to first word |
| EmBench fully solved cases | About 49% for the strongest tested models | Every required criterion satisfied |
| EmBench individual criteria | About 90% satisfied | Partial-credit performance across the requirements |
| Global Work factual accuracy | About 75% with search | Separate 601-question factual benchmark |
A high average can still miss a required step
EmBench contains 295 tasks across eight countries, created with 29 local experts. Its distinction between criteria and complete cases is particularly useful in payroll and employment work. An answer can satisfy most requirements yet miss the exception that determines the employee’s entitlement or the correct calculation. The 49% and 90% figures describe different denominators, not contradictory results.[1]
Keeping the model current is a second engineering task. Deel says about 15% of knowledge-base pages have changed each month since August 2026. It therefore treats memorised expertise as something to version and update, with retrieval checking facts that memory handles less reliably. The company says the models were not trained on customer or personal data.[2][1]
The approach makes a useful distinction between knowing the shape of a problem and verifying its current details. Fast recall can supply the employment concepts, country context and likely procedure. A current source can establish the rule or figure that actually applies. Pangaea’s product value will come from making that combination work within the everyday workflow, while Emet makes the difference between a plausible answer and a complete one easier to measure.
Sources & further reading
- Introducing Pangaea and Emet: AI that knows the rules of global work Deel · 21 Oct 2026 · Accessed 25 Oct 2026
- How We Taught an Open-Weight Model Deel’s Knowledge of 150+ Countries Deel AI Team · 21 Oct 2026 · Accessed 25 Oct 2026