An HR team asking an AI assistant about employee leave needs more than a convincing explanation. It needs the correct rule for the country involved, quickly enough to help someone make a decision before an error reaches a paycheck.
Payroll and HR software company today Deel is betting that smaller AI models, trained to understand that work, can answer many such questions more efficiently than the biggest general-purpose alternatives.
Today, it announced Pangaea, a pair of specialized language models, alongside Emet, a business that supplies expert-created training material and evaluates AI performance.
Together, they give Deel a way to improve its own assistant and sell the expertise behind it to other AI developers.
The clearest early result concerns Pangaea K1, its knowledge model. In Deel's evaluation, K1 scored 70 against a leading score of 75, while completing answers in a median 7.7 seconds. The four comparison models with higher scores required between 14.8 and 28.4 seconds, according to materials Deel supplied to VentureBeat.
Those are company-reported results on a company-developed test. But they illustrate a practical enterprise question: How much should a software vendor spend calling an outside AI model when it can teach a smaller one to handle recurring questions itself?
Two models, built for different jobs
Pangaea starts with K1, which addresses employment rules and operating procedures, and P1, which specializes in payroll.
Deel adapted existing models with downloadable weights rather than training entirely new foundation models. In written answers to VentureBeat, the company said K1 began with an unnamed model containing 9 billion parameters, the numerical values that determine its behavior.
Deel continued training that model using its internal reference material, then taught it to produce answers with examples generated by a more capable teacher model. That process changes the model itself, allowing it to recall information without first retrieving every relevant document.
The company says its reference collection contains more than 43,500 pages maintained by over 2,000 specialists in employment law, human resources, pay administration and regulatory requirements. K1's training material includes rewritten versions of those facts and a limited amount of general text, totaling roughly 1.2 billion tokens.
P1 follows a different path. Deel said it builds on Nvidia's Nemotron family, with AI21 Labs performing additional training using payroll examples produced through Emet. Deel has not identified the precise Nemotron version or K1's underlying model.
Importantly for employers, Deel says neither launch model learns from identifiable employee information or customers' actual payroll records. The training uses documents and examples that specialists prepare.
The upgrade happens inside Deel
Customers will encounter Pangaea through Deel AI, the assistant already embedded in its software. Its intended work includes explaining local employment requirements, leave entitlements and payroll issues.
That does not mean every question will go to Pangaea. Deel keeps general-purpose models for requests that require their capabilities, including broader reasoning and complex agent workflows.
During an interview with VentureBeat, Abhijit Mehta, whom Deel identifies as a senior director of product, described a system that chooses among models according to the question.
“We picked the best model for the best task,” Mehta said.
Customers do not have to make that selection themselves.
“They'll just get the best answer as fast as we can,” Deel cofounder and CEO Alex Bouaziz told VentureBeat.
Deel's separate connection to Claude through the Model Context Protocol, or MCP, remains a different arrangement. Someone querying Deel from Claude still receives Claude's response using the information that connection supplies, the company clarified. Connecting Deel does not automatically switch Claude to Pangaea.
Deel also says document retrieval remains useful alongside K1: it can supply citations, verify details and account for changes since training. Learning employment rules does not eliminate the need to check whether those rules have changed.
What the benchmark actually shows
EmBench Global Work tests 601 questions covering company procedures and international employment topics. Deel gave competing models search access to the same reference collection and used another AI model to grade individual facts in their answers.
Its benchmark chart places K1 above Gemini 3.8 Flash, which scored 69, GPT-6 Sol at 59 and GPT-6 Luna at 52. DeepSeek V4.1 Flash leads at 75, followed by Claude Opus 5.5 at 74, GLM 5.3 Flash at 72 and Claude Sonnet 5.5 at 71.
The comparison therefore does not establish that K1 is the most accurate option. It shows that a specialized model approaches the leaders on this particular evaluation while finishing answers sooner.
Training also changes the outcome substantially. Deel's chart gives the original 9-billion-parameter model a score of 19 without search and 46 with search, compared with K1's 70 after adaptation.
These scores describe performance under Deel's grading method. A score of 70 does not, by itself, mean the system correctly completes 70% of payroll transactions or can independently authorize employment decisions.
The test also compares complete systems with different approaches: K1 recalls learned information, while competitors use retrieval. That reflects Deel's product design, but it is not a universal ranking of the underlying models.
Emet also introduces EmBench Payroll, an evaluation drawing on specialists' knowledge of country-specific pay requirements. Deel says P1 approaches leading models on difficult questions in that test. That is a separate claim from K1's results; neither evaluation establishes reliable performance across every country and payroll scenario.
VentureBeat could not access the public benchmark site during reporting and has not independently reproduced the scores.
Why Deel wants an AI training business
Emet extends beyond checking Deel's own chatbot.
The company describes three offerings: evaluations, expert-produced training examples, and practice environments where AI systems attempt business tasks and receive feedback from automated checks.
The commercial argument rests on Deel's existing infrastructure. Its public company profile lists more than 40,000 customers and operations spanning 150 countries. Deel already recruits, verifies, contracts with and pays specialists across jurisdictions; Emet applies that machinery to people creating and assessing AI training material.
Bouaziz said Deel has begun working with major AI labs, although he did not name them in the interview. He sees business operations as an opportunity to supply expertise that developers need when moving beyond general chat and coding.
Deel acts as the initial proving ground. Experts identify shortcomings, prepare additional examples and measure subsequent versions. Mehta said the team had produced daily internal model revisions during the preceding three weeks and expected that a weekly update schedule might eventually make sense.
That remains an expectation, rather than a promised release timetable.
The new products also complement Akai, Deel's separately available automation platform. Akai coordinates agents performing operational workflows; Pangaea supplies specialist answers; Emet develops the material and tests that can improve model performance. Akai customers do not need a separate Deel payroll or HR subscription.
The enterprise case comes down to cost and control
For Deel, quicker answers offer a customer benefit, while cheaper model calls could improve the economics of delivering AI throughout its software.
Bouaziz emphasized that spending on customer-facing AI affects gross margins. Bringing more specialist requests onto models the company controls could reduce dependence on outside providers and make the cost of serving those requests easier to manage.
That is different from demonstrating lower total ownership costs. Training, hosting, evaluation, document maintenance and expert review still require investment. The materials reviewed do not establish an independently verified monthly savings figure.
Deel says its assistant remains part of existing product subscriptions, without a separate charge for each question. It has not announced downloadable Pangaea weights or a self-service Pangaea API with published usage prices. Emet directs prospective customers toward individually scoped projects rather than a public price list.
For enterprise buyers, the immediate proposition is therefore an upgrade within an existing service. For AI builders, Emet offers a potential source of specialized examples and evaluation expertise.
The larger test will be whether Deel can preserve speed and cost advantages while keeping answers current across jurisdictions. Its own results already show the balance it must strike: specialization makes a smaller model competitive, but some general-purpose systems still answer more accurately.