Deel’s Pangaea models put payroll expertise on a faster track

Two models, two training paths

Deel fine-tuned existing models with downloadable weights rather than training new foundation models from scratch. The company says K1 is based on an unnamed model with 9 billion parameters. It trained the model on internal reference material, then used examples from a more capable teacher model to teach K1 how to produce answers.

Deel’s reference base contains more than 43,500 pages, supported by over 2,000 specialists in employment law, HR, payroll and regulatory requirements. K1’s training material consists of rewritten versions of this information, plus a small amount of general text: about 1.2 billion tokens in total.

P1 follows a different route. According to Deel, it is based on Nvidia’s Nemotron family and was further trained by AI21 Labs on payroll examples prepared through Emet. Deel has not identified the exact Nemotron version or K1’s base model.

The company says neither model was trained on personally identifiable employee data or customers’ real payroll records. Its training relies on documents and examples prepared by specialists.

Deel will offer Pangaea through Deel AI, its assistant for local hiring requirements, leave eligibility and payroll questions. But Pangaea will not handle every query. Deel says it will continue to use general-purpose models when a request calls for broader reasoning or complex agent workflows. Customers do not choose the model: the system selects one based on the task.

A separate connection lets users query Deel from Claude through the Model Context Protocol (MCP). In that setup, Claude still generates the answer using information supplied by the connection; it does not automatically switch to Pangaea.

Document retrieval also remains part of the system. Deel says it can provide sources, help verify details and account for changes made after a model was trained. That matters in employment and payroll, where learned knowledge can go stale.

What Deel’s benchmark shows

EmBench Global Work tests 601 questions about company procedures and international employment rules. Deel gave competing models access to the same reference base, then had another AI model score individual facts in their answers.

K1 scored 70, ahead of Gemini 3.8 Flash at 69, GPT-6 Sol at 59 and GPT-6 Luna at 52. DeepSeek V4.1 Flash led with 75, followed by Claude Opus 5.5 at 74, GLM 5.3 Flash at 72 and Claude Sonnet 5.5 at 71. The four models above K1 took between 14.8 and 28.4 seconds to respond, compared with K1’s median of 7.7 seconds.

Those figures come from Deel’s own test, not an independent evaluation. They also compare whole systems using different approaches: K1 reproduces learned information, while competitors use retrieval. The result says something about Deel’s product design, but it is not a universal ranking of the underlying models.

The training process appears to account for much of K1’s score. Deel’s table gives the original 9-billion-parameter model 19 points without retrieval and 46 with it; the fine-tuned K1 scored 70.

A score of 70 does not mean the system correctly completes 70% of payroll operations, or that it can make hiring decisions on its own. Nor does the benchmark establish that K1 is more accurate than every competitor. It shows that on this particular test, a specialized model came close to the leaders while responding faster.

Emet has also introduced EmBench Payroll, a test based on experts’ knowledge of payroll requirements across countries. Deel says P1 approaches leading models on difficult questions in that benchmark. That is a separate claim from K1’s results, and neither evaluation establishes reliable performance across every country or payroll situation. VentureBeat was unable to access the public benchmark site and independently repeat the testing while preparing its report.

Emet turns internal expertise into a product

Emet is intended to serve customers beyond Deel. Its work includes:

The business draws on Deel’s existing operations. The company’s public profile lists more than 40,000 customers across 150 countries. Deel already hires, vets, contracts and pays specialists in multiple jurisdictions; Emet uses that infrastructure to work with people who create and evaluate AI training material.

Deel co-founder and CEO Alex Bouaziz told VentureBeat that the company already works with large AI labs, but did not name them. He said Deel could supply business-operations expertise to developers moving beyond general chat and coding.

Deel is also the first place to test the approach. Experts identify gaps, prepare additional examples and evaluate updated models. Product senior director Abhijit Mehta said the team had released internal model versions daily for the previous three weeks. He said weekly updates might eventually make sense, but that is an expectation, not a promised schedule.

Pangaea and Emet complement Akai, Deel’s separately available automation platform, which coordinates AI agents carrying out work processes. Pangaea provides expert answers; Emet prepares material and tests that may improve the models. Akai customers do not need a separate Deel subscription for payroll or HR services.

The economics are not yet settled

Deel says the assistant is included in existing product subscriptions, with no separate charge for each question. It has not announced downloadable Pangaea weights or a public API with published prices. Emet offers prospective customers projects with individually scoped work rather than an open price list.

Bouaziz has said AI costs affect gross margin. Routing more specialized questions to models Deel controls could reduce its dependence on external providers and give it more control over the cost of answers. But that is not the same as proving lower total costs: training, hosting, evaluation, document maintenance and expert review all require investment. The materials reviewed do not include an independently verified estimate of monthly savings.

I think the more interesting test is not whether K1 can beat a general model on Deel’s benchmark. It is whether Deel can keep answers fast and current across jurisdictions without spending the savings on retraining, retrieval and human review. Specialization looks promising in the company’s own results, but some general-purpose systems still score higher—and payroll is exactly where a stale answer can erase the value of a fast one.