Deel launched Emet, a new AI lab for model evaluation and training, and Pangaea, a family of language models for global work, on 21 October. Emet published two public payroll and global-work benchmarks alongside them, and Pangaea becomes the main model behind AI at Deel.
What Deel announced
In its launch post, Deel describes Emet as a lab “for model evaluation and training” that specializes in payroll, HR, compliance, legal and finance. Its experts are practicing specialists from Deel’s own teams and its global network, and Deel finds, verifies and pays them. Deel says Emet is open to every model builder.
Pangaea is the first product of that work. Deel describes it as a family of models for hiring, paying and managing people across 150+ countries, built on leading open-weight models. Deel says it trains them on its knowledge base and on expert-built data from Emet, and never on personal or customer data.
What the benchmarks measure
Emet’s public benchmarks are called EmBench, and every model, Deel’s included, is scored against expert answers. Deel published two:
- EmBench Payroll holds 295 real payroll tasks across 8 countries, built with 29 in-country payroll specialists. Deel reports that the best model it tested meets 90% of the specialists’ criteria but fully solves only 49% of the tasks.
- EmBench Global Work holds 601 questions on working across borders, from notice periods to who can legally start a job, scored fact by fact. Deel reports that the best model, with search, gets 75% of the facts right.
Both payroll figures come from the same benchmark. A model can meet most of the criteria on a task and still leave the task unsolved, and the fully-solved rate is the number a payroll team has to care about when a paycheck must be right.
The two Pangaea models
Deel names the first two models after the benchmarks they target.
Pangaea K1 covers global work: notice periods, leave rules and hiring in a new country. Deel trained it on a knowledge base of 43,500+ pages so that it answers from memory instead of searching. On EmBench Global Work, Deel says K1 comes within five points of the top model with search, answering up to 3x faster. Deel adds that training on its knowledge base took K1 from 19% to 70% of facts right when answering from memory.
Pangaea P1 covers payroll. It was trained on payroll data built by Emet’s in-country specialists. On the 110 hardest tasks in EmBench Payroll, Deel says that training moved P1 from 55% to 61% of criteria passed, at about 40x lower cost than the top model.
On speed, Deel says Ask Deel returns a full answer in about 8 seconds, where leading models with search took up to 28 seconds in its testing. Deel also says Pangaea costs a fraction of leading models per answer, which lets it keep AI included in a Deel plan with no charge per question.
How Deel built K1
A four-part technical series from Deel’s AI team describes K1 as a 9-billion-parameter open-weight model. The knowledge base it studied holds about 43,500 pages, roughly 69 million tokens, and about 15% of pages have changed each month since August 2026.
The team found that showing each fact once or twice did not work. It rebuilt the training corpus so every chunk of a page gets around a dozen rewritten “views” aimed at different readers, and the final corpus reached about 1.2 billion tokens, roughly 18 times the size of the knowledge base. Deel reports that continued pretraining raised the model’s score on its evaluation set from about 19 to about 54, and fine-tuning on teacher-written answers raised it to 70.
The series is also direct about the limit of the approach. Deel writes that memory “can be confidently wrong, it cannot show its sources, and it goes stale,” and says it expects Pangaea to pair memory with a light retrieval step for checking specifics, catching recent changes and citing sources.
What Deel says comes next
Deel says Pangaea will keep getting faster and more accurate, and that Emet will add benchmarks across more of how businesses run, starting with payment reconciliation exceptions and sales tasks. It also plans environments where AI agents practice real business tasks from start to finish. The technical series says reasoning and tool calls are in the pipeline but not yet part of Pangaea’s fine-tuning.
What it means for the HR leader
Every number above comes from Deel and from a lab Deel runs. Deel states that Pangaea sits on the same public leaderboards as the other models, so a buyer can check the results, and nothing in the post describes an outside review of the benchmarks.
In our reading, the 49% figure is more useful to an HR or payroll team than the launch itself. It gives buyers a task-level pass rate to ask every vendor for, including Deel, when a vendor says its assistant “knows” payroll.
Other HR suite vendors have also described how their agents will be controlled. We covered Workday’s autopilot agents for hiring and payroll, which come with customer-set limits, and SAP’s autonomous HCM vision, which links HR agents to work redesign. Deel’s launch takes a different route and trains a model on the vendor’s own rules, then publishes the scores.
Three things to do before the next payroll or HR AI renewal:
- Ask each vendor for its fully-solved rate on real payroll tasks in your countries, separate from any partial-credit score.
- Ask which model answers a given question, since Deel says Pangaea becomes the main model, and what happens to answers when the underlying rules change.
- Confirm in writing that employee and payroll records are not used to train any model, using Deel’s own statement on customer data as the template.
Source: Deel, “Introducing Pangaea and Emet: AI that knows the rules of global work”
