Over the past two years, the AI industry's narrative focus has undergone a subtle but profound shift.
In 2024, the core question in market discussions was "whose model is stronger." Parameter counts, leaderboard scores, and benchmark rankings were the main yardsticks for measuring an AI company's value. By 2025, as the capabilities of foundation models gradually converged, the industry's focus shifted to "who can actually put models to use" — AI agents became the new buzzword, with all kinds of coding agents and browser agents emerging one after another.
Now, entering 2026, the question has moved forward again: not "can it be put to use," but "can it deliver real results in actual business operations."
This shift has triggered an explosion of a concept — FDE (Forward Deployed Engineer). Leading companies like Anthropic and OpenAI are recruiting FDEs in large numbers, yet most people still have a vague understanding of what the role entails. An FDE is not pre-sales, not consulting, not after-sales support. Its essence is bridging the gap between platform products and the delivery of final value at the front line — when the connection is made, technology can create value; without it, even the best models produce no results.
Since founding Mininglamp (秒针系统) in 2006, MingLue Technology has been practicing FDE in China. Over twenty years, we have accumulated more than 2,100 clients across industries such as advertising and marketing, retail, and finance, with our core competency always being helping enterprises with data governance and intelligent analytics (Data Intelligence).
Starting in 2025, with the rapid development of the company's Agentic AI technology, our business structure underwent a critical reorganization: on top of our existing data intelligence services, a brand-new category emerged — Agentic Service. The difference between the two can be summed up in one sentence: Data Intelligence helps clients "see the data," while Agentic Service helps clients "do things with the data."
Behind this lies a new model in which an FDE leads a group of AI agents working collaboratively through the Octo platform. Currently, more than 1,400 MingLue employees and over 3,200 AI agents collaborate daily on this platform.
At a recent analyst briefing, MingLue Technology founder, CEO, and CTO Wu Minghui gave a systematic presentation on "FDE in AI Native organizations."
Below are the minutes of the analyst briefing.
Q1: Please systematically introduce the origins of FDE and how it has changed in the AI Native era.
Wu Minghui: FDE (Forward Deployed Engineer) is a concept pioneered by Palantir, referring to Forward Deployed Engineers. When we established the MingLue brand in 2014, its English name Mininglamp (Magic Lamp for Data Mining) was inspired by Palantir, so we have indeed done a lot of research and thinking on FDE.
FDE is not a concept that became popular only after the advent of artificial intelligence. It has existed since Palantir's early days of big data analytics, when it sent its best engineers to work on-site at clients. Traditional software outsourcing or delivery companies lack the ability to improve the product after delivery. But Palantir's front-line teams not only served front-line clients, they also brought the know-how back to improve their own middle-platform products — most importantly data governance, that is, the construction of the ontology.
A core reason behind this is "Context, not Control." This concept comes from the management philosophy of Netflix founder Hastings: fully delegate the context of all information to front-line teams, letting them make their own decisions once they have the information, forming collective wisdom and emergence from the bottom up. Palantir also drew on this idea in military scenarios. The collaborative office products of internet companies are, in essence, also about maximizing shared Context within an organization.
So the essence of FDE is bridging the enormous gap between an enterprise's IT middle platform and front-line business staff. If this gap is not bridged, neither AI capabilities nor big data capabilities can create value.
But in the AI Native era, the FDE model has changed. In the big data era, FDEs mainly did client-side data governance work, helping front-line business staff build data middle platforms and sharing context with people. Today, every front-line business person has their own AI agent, so context must be shared not only with people but also with AI. The AI middle platform must not only share context, but also share skills, share compute, and share model gateways.
So today, the essence of FDE is forming a shared objective with the client, obtaining enough context from the client, and feeding the client's taste back to the AI through continuous collaboration. These three things are all things AI does not know. Once they are obtained, AI can produce the right output.
For future FDEs, reasoning and thinking abilities become secondary, because these can be taken over by AI. What matters more is the ability to build trust with clients and communicate and interact with them. The FDE is more like a salesperson, bringing back what the client truly wants and the client's taste, and then AI can help produce the final result.
Q2: What are the differences between MingLue's Octo platform and the previous generation of online collaboration platforms? Why did you choose open source?
Wu Minghui: Today, AI's coding ability has become powerful enough to redefine the development efficiency of large-scale software. A collaboration platform that once took thousands of people years to develop can now have its core features built well by a few dozen engineers working with AI. The first version of Octo was developed by fewer than 10 R&D colleagues, and many users who downloaded and deployed it from GitHub gave very positive feedback.
We chose open source because large enterprises all want an open-source platform to protect their Context. In the AI era, this Context data is a core enterprise asset. The truly valuable production data resides in conversations between people and agents, but the previous generation of online collaboration platforms does not yet support one-click export.
At the same time, many hardware vendors also want to connect to an open AI collaboration ecosystem. Octo's open source provides exactly such a foundation — just as Nokia and Xiaomi embraced Android back then and together built a rich ecosystem of smart devices. We believe collaboration platforms in the AI era will follow the same path toward openness.
Q3: What is the core driving force behind MingLue's transformation into an AI Native company? Is your relationship with large model companies competition or cooperation?
Wu Minghui: MingLue has been doing AI since the day it was founded — we are not a company that transformed into AI Native. Over the past twenty years, our Data Intelligence has always been accumulating data for AI. As early as 2020, we began investing in AI agent R&D. After AI's coding capabilities broke through, we seized the opportunity to launch quickly, and what we had long wanted to do finally came at just the right time.
Today we have chosen a proprietary model strategy. The opportunity we see is this: foundation models have general capabilities, but when entering business scenarios, clients must consider cost-effectiveness. At the same time, clients need to securely connect their private data with model capabilities — requiring both model training capability and the ability to connect to clients' private data. Clients certainly need smart models, but they need security, reliability, and controllable costs even more — and that is MingLue's opportunity. The leading results achieved by our self-developed models such as Mano-P on authoritative global leaderboards provide strong technical support for serving our clients.
Q4: What challenges do clients commonly face when transforming into AI Native? What is MingLue's advice?
Wu Minghui: All enterprises are now very concerned about this issue. We've found that when providing customers with AI Native system construction, it's not simply a matter of deploying Octo. We also need to equip customers with a group of "consulting agents" that embed transformation frameworks like Kotter's Eight-Step Model, helping the CEO first think through how to operate, and then design the transformation path.
MingLue's inherent advantage is that we have a large number of AI Native businesses in progress. We use AI to transform professional teams and distill methodologies into agents, similar to the collaboration model between AI companies like OpenAI and Anthropic and Blackstone or Goldman Sachs. In the future, we won't need to station large numbers of consulting engineers at customers—one FDE leading a group of agents can complete the delivery of AI transformation.
The most critical point: the top leader must personally get involved. For enterprises still in the exploration stage, we hope to help them shorten this process through the Octo platform and consulting agents.
Q5: Once enterprises have AI capabilities, what long-term value does MingLue provide?
Wu Minghui: Currently, the AI Native transformation of enterprises requires systematic engineering. Our core growth driver is Agentic Service. In vertical scenarios such as advertising and marketing, we use AI agent teams to directly deliver business results.
In the long run, our business model is a three-layer stack:
The first layer is pay-for-performance. Customers pay at the price standards of traditional service providers, and we take a revenue share based on leads or GMV.
The second layer is custom development services. After enterprises adopt the AI Native platform, a large number of business scenarios require developing dedicated agents. The previous generation of SaaS was software for people to use; in the future, one scenario equals one agent. Our FDEs, together with agents, help customers complete development, charging service fees based on token consumption. At the same time, we are post-training our own code models to deliver at lower cost and improve the gross margin of this business.
The third layer is private post-training services. After large enterprises build their own computing power, we help customers complete model post-training and deployment on their own infrastructure. Many enterprises need to optimize token consumption and reduce long-term usage costs—this is exactly where we can continuously provide value.
These three layers are not substitutes for one another; they unfold gradually as customers' AI maturity increases. Even when enterprises have basic AI capabilities, they still need professional partners in areas such as industry know-how, proprietary models, and engineering implementation.
Q6: There is a view in the market that large enterprises' token bills are spiraling out of control. How does MingLue help enterprises convert tokens into effective value?
Wu Minghui: We have observed that agentic tasks are driving rapid growth in token consumption at scale. The essence of Agentic Service is decomposing complex tasks into subtasks—a complete agency-operation task may consume thousands of times more tokens than ordinary conversational interaction. But some enterprises indeed cannot afford it, mainly because frontline employees are using tokens to handle low-complexity tasks rather than driving business processes that directly generate revenue.
Our solution is to help customers use tokens rationally to produce higher value. In Agentic Service scenarios, we don't settle by tokens but take a revenue share based on business results. FDEs can see revenue and token costs themselves and calculate P&L—how much business result each token corresponds to is visible. In custom development scenarios, the development workload of agents is itself quantifiable, and we charge service fees by tokens, letting customers pay for the computing power actually consumed.
At the same time, we believe enterprises also need to use models in tiers. High-frequency, low-value tasks should use dedicated small models, while high-value tasks should use more capable models—that is the reasonable state.
About MingLue Technology
MingLue Technology (2718.HK), founded in 2006, is China's leading Agentic Service enterprise with proprietary model capabilities. In 2025, it listed on the Hong Kong Stock Exchange as the "world's first Agentic AI stock." It has twice won the Wu Wenjun Artificial Intelligence Science and Technology Award, has been repeatedly featured in reports by Gartner and IDC, and holds more than 2,400 technology patents and over 500 software copyrights.
In recent years, MingLue Technology has continued to make breakthroughs in the Agentic AI field: in 2024, its self-developed hypergraph multimodal large model (HMLLM) technology achievement received a Best Paper nomination at the global top conference ACMMM 2024; in 2025, it fully launched the DeepMiner proprietary large model product line, with its VLA model Mano topping both the Mind2Web and OSWorld global SOTA leaderboards; in 2026, it open-sourced the on-device GUI-VLA agent model Mano-P, topping nine leaderboards including OSWorld, ScreenSpot, and MMBench, ranking first on the OSWorld dedicated-model leaderboard; it open-sourced the Apple Silicon inference SDK Cider, accelerating on-device inference by up to about 1.9x; and it open-sourced the agent collaboration platform Octo, building an agent collaboration network for the IOA era, and launched its first AI Native recording hardware, Octic, extending AI agent capabilities from platform to hardware and building an Agentic AI "model-platform-hardware" closed loop.
Backed by 20 years of technological accumulation, MingLue Technology has served 135 Fortune Global 500 companies, approximately 2,100 brand customers, and more than 240,000 enterprise users, covering industries such as retail, consumer goods, automotive, and 3C.