While DingTalk's previous support for the AI ecosystem centered around foundational assistance in resources and commerce, its commitment has now evolved to a more profound level—genuine AI model training capabilities. This enhancement offers comprehensive support to entrepreneurs developing vertical large models, significantly enhancing startup success rates and the certainty of model product commercialization.

Author | Mr. Pi

Produced by | Industrialist

90.2%—this remarkable figure represents the performance of EisonHealth's 'Doukou Gynecological Large Model' in professional testing.

On June 30, this Hangzhou-based medical startup successfully completed 100 gynecological professional test questions using a large model trained on DingTalk's enterprise-exclusive AI platform. The test questions covered six major gynecological symptoms, including irregular menstruation, abnormal bleeding, abnormal vaginal discharge, lower abdominal pain, and lower abdominal masses. The results were highly consistent with professional doctors' diagnoses.

Achieving this feat was no easy task. From a professional perspective, compared to the training of base large models, industry-specific models often struggle to ensure high accuracy due to the scarcity of high-quality data and high training costs. The Doukou Gynecological Large Model, for instance, reportedly had an accuracy rate of only 77% in April.

DingTalk represented a new turning point.

More specifically, over the past month, with DingTalk's support, the EisonHealth team made various adjustments in data processing, computational power, model optimization, and other areas. As a result, the model's accuracy rate soared from 77% to over 90%, marking a more than 15% improvement.

This development signifies another positive step for DingTalk in supporting AI ecosystem enterprises. Earlier this year, in March, DingTalk introduced a series of ecosystem policies for entrepreneurs, including startup funds, dealer systems, and other forms of support. Now, four months later, DingTalk has taken another stride towards the ecosystem's underlying technology by providing deeper model product capabilities for its ecological partners.

What kind of AI ecosystem does DingTalk aspire to build? Or, in the current wave of AI startups, what is the value of DingTalk as an ecological foundation? These questions are gradually being answered with the launch of the Doukou Gynecological Large Model.

I. In the Wave of Vertical Large Models, an Emerging 'True Bottleneck'

'Improving the accuracy rate of industry large models from 77.1 points to 90 points is a significant achievement,' said Zhu Hong, CTO of DingTalk. 'In our collaborations with many partners, DingTalk has discovered that building, deploying, and applying exclusive large models present challenges from the initial stages of demand positioning and data preparation. Often, teams do not know where to start and cannot guarantee the final results.'

This accurately depicts the current landscape.

If there's a prevailing theme this year, it's undoubtedly the deployment of large models. Among these, the deployment of industry-specific vertical large models and corresponding scenario agents is even more crucial.

According to data from the '2025-2030 China AI Large Model Industry Competition Pattern Analysis and Future Trend Prediction Report' published by China Research and Intelligence, China's AI large model market size is expected to exceed 49.5 billion yuan by 2025, with vertical models accounting for over 60% of the market space. Similarly, a set of forecast data from the Co-Research Institute predicts that by 2028, the market size of vertical models will reach 62.4 billion yuan.

However, this journey is not without obstacles. According to incomplete statistics, among the current vertical large models on the market, most models' accuracy rates are concentrated in the 60%-70% range. The more segmented the scenario, the lower the accuracy rate of the model.

The reasons are straightforward. As Zhu Hong explained, for most startups, providing customers with sufficiently adaptable AI vertical model services often entails practical difficulties in model construction.

For example, on the demand closed-loop side, although developers and entrepreneurs can clearly identify targeted needs and scenarios, most enterprises or entrepreneurs lack the ability to further dissect the specific 'scenario to model' link. This makes it challenging to conduct model training and tuning based on a specific scenario link closed loop.

Another challenge lies in data. For some enterprises or developers, despite accumulating data over the years, transforming industry data into high-quality data that can provide positive feedback to the model remains an engineering and experience challenge for model training.

On the AI solution side, even after training the model, most developers and entrepreneurs may not have the capability to transform the model product into an application product or construct corresponding industry solutions based on other supporting services.

Moreover, there are inevitable costs. For most startups, conducting model training locally can be expensive. Taking DeepSeek as an example, the deployment cost for the full-featured version is often at least several million yuan. If one aims to reduce the cost to within one million yuan, the corresponding accuracy of the model will decrease, and the training and inference effects will not be optimal.

Issues such as optimizing model training methods, the difficulty of building a training talent pipeline, data security, and commercialization are also practical problems faced by many entrepreneurs building vertical large models.

Where lies the solution? Or where is the ideal entrepreneurial soil for entrepreneurs navigating vertical industry models?

II. DingTalk Ecosystem Takes Another Step Towards AI

The successful deployment of the Doukou Large Model underscores DingTalk's significant progress in the AI ecosystem.

While DingTalk's previous support for AI startups focused more on funding and the commercial environment, the platform is now personally stepping into the field to provide more granular product-level support to entrepreneurs.

Taking the Doukou Gynecological Large Model as an example, its performance improvement was achieved through training on DingTalk's enterprise-exclusive AI platform. In the initial training phase, due to limitations in data preprocessing experience, insufficient computational power, and model tuning capabilities, the EisonHealth model training team hit a bottleneck with an accuracy rate of 77.1%. This result was insufficient for practical applications in medical AI.

DingTalk became their choice for the second round of training. With DingTalk's support, the team optimized the model training strategy and overcame limitations in computational power and data. With advancements in specific training methods and service support, EisonHealth swiftly achieved a substantial improvement in training efficiency, reducing the single training duration from 26 hours to 7 hours. Ultimately, the team achieved a leap in accuracy, rising from 77 points to 90 points in a shorter period.

'This is a significant progress. It's akin to transforming a generalist who knows a bit of everything into a specialist comparable to an expert in a short span. This involves effective and secure data preprocessing, efficient scheduling of computational power, the construction of a model evaluation mechanism, and fine-tuning of training algorithms and model parameters,' said Zhu Hong.

In fact, to a certain extent, DingTalk provided a systematic AI support system during the training process of this gynecological large model. This system encompasses not only guidance on model training methods and strategies but also comprehensive support for the entire lifecycle of model development, from model training to model testing and evaluation.

It is precisely based on this support that EisonHealth was able to achieve a fundamental leap, transforming the model from unusable to highly usable and user-friendly.

This is also a microcosm of the current DingTalk ecosystem. To better assist enterprises in training and deploying their professional/exclusive large models, DingTalk has built a set of industry/enterprise large model construction support systems leveraging its inherent product and technical capabilities.

Specifically, enterprises or partners on DingTalk receive full-process platform product support and service assistance, including data collection, cleaning, annotation, basic model selection, model training, effect evaluation, model tuning, and model engineering deployment. This holistic support helps ecological enterprises develop exclusive industry large models more efficiently and deploy related AI applications.

Moreover, the support of this new system extends beyond the product level to solutions and talent. It is understood that in addition to the above assistance, DingTalk will also provide ecological partners with AI solution consulting, industry large model solution consulting, and AI talent training and assessment services. This not only helps enterprises refine usable and implementable products but also provides the necessary soil, fertilizer, and sunlight for the actual landing of model products.

III. The Ecosystem for AI Startups: DingTalk Provides a Mature Answer

A undeniable fact is that with the advent of the wave of AI large models, a new wave of entrepreneurship is emerging. This includes not only newly emerging AI native application entrepreneurs but also seasoned SaaS software practitioners and veterans of smart hardware startups.

However, this entrepreneurship is fraught with a high failure rate. According to data from CB Insights, the failure rate of global AI startups has hovered around 65%-75% over the past three years. This means that out of every three AI companies established, two fail to survive beyond three years. In the Chinese market, this failure rate is even higher. According to data from IT Orange, the failure rate of startups in the entire AI market is as high as 70%-80%.

DingTalk's ecological actions are a further breakdown based on this problematic landscape.

That is, if DingTalk's previous support for the AI ecosystem was more focused on foundational assistance in resources and commerce, its current support has evolved to a deeper level—genuine AI capability support. This provides comprehensive assistance to vertical large model entrepreneurs, comprehensively improving startup success rates and the certainty of model product commercialization.

To some extent, this is precisely the entrepreneurial soil needed by current vertical large model enterprises. Leveraging the DingTalk system, these enterprises can navigate the entire process from industry understanding to AI entrepreneurship. Through DingTalk's large model platform support, they can transform their inherent accumulation in specific industrial directions into usable and implementable model products. Coupled with DingTalk's robust support in talents, suppliers, and other aspects for the landing and implementation of AI products, these enterprises can truly complete their entrepreneurial journey from 0 to 1.

In fact, the potential extends beyond this. After the model product is formed, as entrepreneurs deploy the model product in various industrial scenarios, they can further list it on DingTalk's AI open market, making it accessible to 23 million enterprises. This allows vertical large model products to receive greater exposure and more direct market validation. Simultaneously, based on these validations, a closed loop for model optimization and training can be further constructed.

Taking EisonHealth as an example, the 'Doukou Gynecological Large Model', a large model product realized with the support of the DingTalk AI platform, will next empower women to enjoy more efficient and professional health services through the 'Girlfriend Doctor' product, alleviating the scarcity of gynecological medical resources. Additionally, the gynecological AI agent can be integrated into the DingTalk agent application market to serve more medical beauty institutions on DingTalk, helping them better serve their customers.

It can be said that DingTalk has now constructed a sufficiently complete and self-contained ecosystem. With DingTalk's completion of this more granular product puzzle, we can witness a holistic AI entrepreneurial system within the entrepreneurial ecosystem. This system ranges from model product support to product design and implementation, from funding support to channel partner assistance, and more directly faces the end-customer market.

Currently, whether it's the number of enterprise users, current AI usage, AI component capabilities, AI model compatibility, or a deeper understanding of different industrial scenarios across thousands of industries in China, DingTalk stands out as one of the most prominent names in the current AI wave.

Today, DingTalk is encapsulating these strengths into individual ecological components and opening them up as the platform value of infrastructure. This helps a vast number of small and medium-sized enterprises find new anchors in the surging AI wave, thereby promoting the digital and intelligent upgrading of industrial AI across thousands of industries in China.

'Teach a man to fish, and you feed him for a lifetime.' This is the value of DingTalk and the latest mission of China's largest TO B enterprise platform.