Deel推出HR与薪酬专用AI模型 应答成本最高降98%

Summary 1: Product and cost data

1. Deel has launched Pangaea, a suite of domain-specific large models that includes K1 and P1. K1 handles employment-rule operations, while P1 focuses on payroll calculations; both are fine-tuned from open-source models rather than trained from scratch.

2. K1 uses a 9-billion-parameter base model. Its training corpus includes more than 43,500 pages maintained by over 2,000 experts, or about 1.2 billion tokens after rewriting. P1 is based on NVIDIA Nemotron and was additionally trained by AI21 Labs using Emet cases.

Summary 2: Usage and caveats

1. Pangaea is embedded in the Deel AI assistant, which automatically routes questions to the appropriate model. Complex reasoning still uses general-purpose large models, so users do not need to choose models themselves.

2. K1's per-response cost is about 1/49 of the comparison solution, with a reduction of up to 98%. However, Deel did not disclose the comparison baseline or total cost of ownership.

3. Assistant features are included in existing subscription entitlements and are not charged per question. There are currently no model-weight downloads or publicly priced self-service APIs.

4. K1 scored 70 on EmBench, with a median response time of 7.7 seconds. The results have not been independently reproduced and do not mean the model can independently complete payroll transactions or employment decisions.

Summary 1: Brand positioning and product development

1. Deel uses Pangaea to establish a dedicated AI brand for HR and payroll. K1 focuses on employment rules and operations, while P1 specializes in payroll calculations. Both adapt and fine-tune open-source models and emphasize expert data and domain expertise.

2. Through embedded service in the Deel AI assistant, the system automatically routes questions to the appropriate model, while complex agentic workflows still use general-purpose large models. Document retrieval is retained for citation verification and rule synchronization.

Summary 2: Channels, pricing, and user communication

1. The service reaches Deel's existing 40,000 enterprise customers across 150 countries. The AI assistant is included in existing subscription entitlements with no per-question fee, lowering the barrier to trial. Deel has not yet launched downloadable weights or a publicly priced self-service API.

2. The model context protocol cooperation with Claude is unaffected; Deel queries on the Claude side will not automatically switch to Pangaea. Emet is customized on a per-project basis, with no public price list yet.

3. Marketing can cite data such as the 98% cost reduction and 7.7-second response time, but must note that the benchmark has not been independently reproduced and total cost of ownership has not been calculated, avoiding overpromising.

Summary 1: Market opportunity and positive impact

1. Deel has launched Pangaea and Emet, cutting the cost of specialized HR and payroll responses by up to 98%. K1's per-response cost is about 1/49 of the comparison solution, helping shift more requests to Deel's own controllable models.

2. Deel has 40,000 enterprise customers across 150 countries. The AI assistant is included in existing subscriptions with no per-question fee. Emet externally provides performance evaluation, expert training cases, and hands-on practice environments, creating project-based cooperation opportunities.

Summary 2: Risks and lessons

1. The public benchmark test site is inaccessible, and the scores have not been independently reproduced. They do not mean the system can independently complete payroll transactions or employment decisions. The cost comparison baseline and total cost of ownership were not disclosed.

2. There are currently no Pangaea weight downloads or publicly priced self-service APIs, and Emet has no public price list; cooperation requires customized integration.

3. Deel's approach offers lessons in using expert data, open-source fine-tuning, teacher-model guidance, and evaluation-feedback iteration. It also retains document retrieval and Claude integration to avoid migration shock. Akai users can use the collaboration capabilities without additional payroll or HR subscriptions.

Summary 1: Product production and design requirements

1. Pangaea was not trained from scratch. K1 is fine-tuned from an undisclosed 9-billion-parameter base model, with training materials including more than 43,500 pages maintained by over 2,000 experts, or about 1.2 billion tokens after rewriting. P1 is based on NVIDIA Nemotron and was additionally trained by AI21 Labs using Emet cases.

2. Training does not touch identifiable employee personal information or real client payroll records; the materials are specially compiled by experts. K1 retains document retrieval for citation verification and synchronization of new rules.

Summary 2: Business opportunities and digitalization lessons

1. Emet can externally provide AI performance evaluation, expert training cases, and hands-on practice environments, reusing Deel's cross-jurisdiction expert recruitment, verification, contracting, and payment system to form a training-material production and evaluation business.

2. Under the collaboration with Akai, Akai coordinates and executes operational workflow agents; Pangaea outputs specialized responses; and Emet provides training and testing. Akai users do not need additional payroll or HR subscriptions.

3. On digitalization, the lessons include automatic model routing, delegating complex requests to general-purpose models, and using proprietary models to reduce costs. Risks include undisclosed base-model sources and P1 versions, tests not independently reproduced, and total costs not calculated.

Summary 1: Industry trends and new technologies

1. A domain-specific model path is emerging in HR and payroll: Deel has released Pangaea, with K1 handling employment-rule operations and P1 specializing in payroll. The models are fine-tuned from open-source models rather than trained from scratch, and use teacher-model guidance and direct memorization of information to reduce full retrieval on every query.

2. Deel adopts hybrid routing: the Deel AI assistant automatically routes questions to the appropriate model, while complex reasoning and agentic workflows still call general-purpose large models. K1 retains document retrieval for citation verification and rule updates.

Summary 2: Customer pain points and solutions

1. Pain points include high costs for specialized responses, complex cross-jurisdiction compliance, difficult payroll calculations, and users not wanting to select models themselves. Deel claims K1's per-response cost is about 1/49 of the comparison solution, with a reduction of up to 98% and a median response time of 7.7 seconds.

2. Emet provides AI performance evaluation, expert training cases, and hands-on practice environments, reusing the expert operations system built across 40,000 enterprises and 150 countries. Akai handles workflow execution coordination.

3. Risk note: EmBench scores have not been independently reproduced, and the public benchmark site is inaccessible. The cost comparison and total cost of ownership were not disclosed, and there is no public API or model weights yet.

Summary 1: Platform requirements and latest practices

1. Customers need fast, expert answers on local employment, leave entitlements, and payroll questions, and do not want to choose models themselves. Deel embeds Pangaea in the Deel AI assistant, which automatically routes by question type, while complex reasoning and agentic workflows are still handled by general-purpose large models.

2. The platform retains K1 document retrieval for citation verification and rule synchronization. The model context protocol cooperation with Claude is unaffected, and queries on the Claude side will not automatically switch to Pangaea.

Summary 2: Operations, partnerships, and risk avoidance

1. Operationally, Deel has made daily internal version updates over the past three weeks and may shift to weekly updates later. Emet is customized on a per-project basis and has cooperated with multiple leading AI labs, whose names were not disclosed. Akai users do not need additional payroll or HR subscriptions.

2. Commercially, the AI assistant is included in existing subscription entitlements with no per-question fee. There is currently no publicly priced self-service API, downloadable weights, or Emet price list.

3. Risk communication must be candid: the 98% cost reduction was disclosed without a comparison baseline, and total costs including training, hosting, and evaluation have not been calculated. Tests have not been independently reproduced, so Deel cannot claim the system can independently complete payroll transactions or employment decisions.

Summary 1: Industry developments and business model

1. Deel has released Pangaea, a domain-specific model for HR and payroll that includes K1 and P1, both fine-tuned from open-source models. K1 was trained on more than 43,500 pages and about 1.2 billion tokens maintained by over 2,000 experts. P1 is based on NVIDIA Nemotron and was trained by AI21 Labs using Emet cases.

2. The business model combines an embedded AI assistant with project-based Emet services: the assistant is included in existing subscriptions with no per-question fee, while Emet externally provides evaluation, expert cases, and hands-on environments. There is no public API, model weights, or price list yet.

Summary 2: Emerging issues and policy implications

1. Evaluation issues stand out: EmBench Global Work covers 601 questions. K1 scored 70, behind DeepSeek V4.1 Flash's 75, among others, but had a median response time of 7.7 seconds. The public benchmark site is inaccessible, and the scores have not been independently reproduced.

2. Cost estimates claim a reduction of up to 98% and a per-response cost of about 1/49 of the comparison solution, but the comparison baseline and total cost of ownership were not disclosed.

3. On governance, training does not touch identifiable employee personal information or real payroll records. Test scores do not mean the system can independently complete payroll transactions or autonomously make employment decisions. Cross-jurisdiction compliance and expert qualification systems are important research samples.