Haimayun's Dang Jinfeng: Anyone Can Buy Models; What Sets Players Apart Is an AI Productivity System

On one side, Silicon Valley and major labs are locked in fierce competition on benchmark leaderboards—today claiming parameter counts have doubled, tomorrow boasting that their multimodal metrics are god-tier; on the other side lies an extremely awkward commercialization scissors gap: the vast majority of consumer-facing AI applications are mired in the quagmire of "the bigger the scale, the higher the losses."

As the capability gap between models narrows day by day, the business model of simply wrapping and reskinning APIs has been proven incapable of forming a moat. Alex Mashrabov, founder of overseas AI video dark horse Higgsfield, once asserted: "AI applications are dying from the technical vanity of self-developed models; controlling traffic routing is the real moat."

But this is only half right. Model routing solves "invocation efficiency," but fails to solve "delivery capability." Merely shedding vanity benchmark scores and doing traffic distribution at the front end still cannot escape the fate of being steamrolled by underlying general-purpose models.

Facing this crisis that is like a frog being boiled in warm water, Dang Jinfeng, founder and CEO of HaiMaYun, offered another penetrating industry insight:

"Consumer-grade AI is very hard to make money from; looking globally, only a tiny number of giants can succeed, and the only scenarios that can truly land and monetize are vertical productivity scenarios. The essence of AI's next stage is not divergent probabilistic generation, but deterministic system delivery. Anyone can buy a model; what truly creates the gap is the 'productivity engine' in which compute, the Harness (agents, context, toolchains) and data are deeply interlocked. AI without industry agents and expert trajectories is ultimately just a toy that spits out words."

RunningHub, under HaiMaYun, is precisely the concrete implementation of this engine system in the AIGC field.

SimilarWeb data shows that from March to September this year, total site visits to RunningHub at home and abroad grew by 266%, ranking among the top mainstream AI creation platforms of the same period in growth rate; in July this year, Wen Jian, co-founder and CTO of HaiMaYun, also revealed in an interview with CCTV's "Focus Interview" that the platform's monthly Token consumption had reached 9 trillion.

Breaking down HaiMaYun's technical path, it is not hard to see that this technology platform, which started out in real-time cloud rendering, is demonstrating another solution: compared with front-end applications that simply pile up APIs, a vertical productivity system integrating compute, toolchains and data is the core fulcrum for breaking through the homogenization dilemma at the application layer.

At the AI application layer, discussing gross margin and commercialization while setting aside compute is often building castles in the air. At present, the vast majority of pure application-layer companies on the market (including the Higgsfields) essentially operate on an "asset-light" basis: lacking a self-developed compute foundation, they rely heavily on renting third-party cloud GPUs or directly passing through closed-source APIs.

This directly leads to two fatal consequences: first, gross margins are highly vulnerable to being squeezed at will by upstream compute costs; second, in the face of explosive commercial invocations, compute costs cannot be physically squeezed down.

However, in the real content production workflow for audio, video and images, AI inference and traditional graphics rendering are production workloads that occur alternately. 3D scenes need rendering; professional software such as Blender and Unreal depend on GPUs when running; and complex video generation and super-resolution compositing are extremely compute-intensive. The traditional approach is to stockpile cards separately for inference and rendering according to business peaks, and the result is inevitably "queuing at peaks, cards sitting idle at troughs."

"So what determines cost is not just whether the model is expensive; Tokens are only part of AI production cost. For image, video and 3D production, what truly determines unit delivery cost is whether the entire compute chain from inference to rendering can operate efficiently. What AI production platforms will compete on in the future is not just compute scale, but the efficiency of large-scale organization and operational scheduling of heterogeneous compute."

This is a judgment that application companies without control over underlying GPUs cannot make, and it is also HaiMaYun's natural genetic moat.

HaiMaYun was founded in 2014 and started out in real-time cloud rendering. Behind the cloud gaming services for games such as Genshin Impact and Black Myth: Wukong, its technical support is present; relying on GPU compute scheduling capabilities accumulated over many years, it is also one of the earlier companies in China to direct compute toward AIGC, and has now completed its identity leap from a cloud rendering service provider to an AI productivity platform.

After a decade of accumulation in the real-time cloud rendering field, HaiMaYun has deployed 69 computing nodes worldwide and possesses a 6 EFLOPS compute pool. HaiMaYun's underlying logic is to pool heterogeneous GPUs from the root up, and directly move the "deterministic delivery and millisecond-level scheduling" capabilities accumulated over ten years in the real-time cloud gaming field onto AI inference.

On the RunningHub platform, the same batch of GPUs serves different workloads such as image generation, video generation and professional rendering at different times, with a unified scheduling layer allocating in real time according to task type, model size and latency requirements. This "integrated rendering and inference" compute scheduling pushes the overall compute utilization rate of the HaiMaYun platform to an astonishing level 70% higher than common industry levels.

For paying enterprise customers, this means extremely high cost-effectiveness; and for HaiMaYun itself, the combination of self-controlled compute and vertically post-trained models allows its comprehensive gross margin to stand stably at the high end of the industry. In the AI melee where gross margin is a matter of life and death, this compute-level pooling capability constitutes a cost red line that pure application companies cannot cross.

Instead of competing on general model parameters, it focuses relentlessly on closing the real productivity gaps and bridging the last mile from models to production.

As multimodal large models have developed to this day, one increasingly obvious contradiction is this: the probabilistic generation capabilities of large models are becoming more and more impressive, but once they enter real industrial-grade production pipelines, a huge gap still yawns between the results and the expectations for commercial delivery.

The reason behind this, according to Dang Jinfeng, is very straightforward: in the field of content production, divergent creative demands mean that general models will always have shortcomings.

Most tasks handled by large language models are convergent: whether code runs, whether a math problem is solved correctly, there are clear benchmarks to measure them; but images, audio, and video are divergent. Aesthetics, moreover, has no standard answer. Once divergence lands in rigorous industrial production environments, it instantly exposes itself as concrete engineering defects—character face changes, style drift, lighting and shadow errors, out-of-control camera work.

A deeper problem is that creators' highly individualized aesthetic expression preferences, as well as all kinds of wildly imaginative special effects that fall outside the distribution of conventional pretraining data, simply cannot be accurately understood and stably generated by a general foundation model from just a short prompt.

General models pursue broad applicability, while real film-grade production requires pixel-level precise controllability. Between these two goals, there is an inherent capability gap.

Based on this insight, HaiMaYun made a strategic choice: give up competing with big companies on the parameter scale of general models, and instead concentrate all core resources on scenario-specific post-training and engineering optimization of general models.

Relying on its heterogeneous GPU compute scheduling foundation and deep engineering capabilities accumulated over many years, HaiMaYun relentlessly pursues scenario implementation along two paths, bridging the last mile from algorithms to production:

First, vertical model post-training to solve the "hard engineering problems" of enterprise-grade delivery: targeting the most frustrating constraints in real production—speed, cost, quality, and consistency—it launched the RH series of vertical models.

In just 20 days in September this year, HaiMaYun launched 3 exclusive RH series scenario-specific post-trained models:

△Among the 13 tested configurations, RH Upscale Ultra ranked first in overall quality with a score of 0.539; Max completed 2x super-resolution of a 5-10 second test video in an average of 34 seconds.

This ability to relentlessly pursue scenario implementation has even been recognized by the open-source "original manufacturer"—MiniMax officially is procuring H3 RH Enhanced in reverse through RunningHub for its customers to use. A third-party model that has undergone scenario-specific post-training being output back to a large model vendor is extremely rare in the current AI industry. This means that the decision-making power for model invocation is shifting to feedback from the production floor.

Second, opening up the fine-tuning ecosystem mechanism and turning "unquantifiable aesthetic preferences" into assets: in addition to the platform solving general problems, HaiMaYun has also, through practice over the past two years, built up a highly defensible fine-tuning mechanism. Facing individual unique aesthetics, specific art styles, and extremely high consistency demands that general models cannot understand, HaiMaYun supports creators and enterprises in repeatedly training and debugging with their own works, generating exclusive Checkpoint fine-tuned models and stylized LoRAs, which are then seamlessly integrated into automated workflows for long-term reuse.

This mechanism precipitates originally abstract aesthetic flavors into "digital art assets" that can be precisely executed by AI. At present, a massive number of such style LoRAs have accumulated on RunningHub, and even a leading mobile game company has trained its entire game visual system into an exclusive model matrix, achieving industrial-grade standard output from concept art to scenes.

Dang Jinfeng said frankly: "We do not train models for the sake of owning models; we train them only so that AI can deliver results in real production. The AI industry does not need every company to build large models, but it does need someone to make large models truly do the work well. What HaiMaYun focuses on is not how big the model is, but how far the model still is from real production. The endpoint of model competition is not parameter scale, but production capability."

Many pure application companies at home and abroad remain at the model invocation layer: they make an interface for the model, wrap it in a chat box or canvas, and let users drive AI through prompts. But production is not chatting, and real tasks cannot be completed with a single prompt. They require task decomposition, multi-model collaboration, tool invocation, process validation, and continuous iteration, ultimately delivering results that meet requirements.

The chat box solves how to use a model, while the production system solves how to make the model complete work on the production line. This system responsible for organizing models, context, and the production toolchain, and for being accountable for final delivery, is Harness. It determines whether AI can move from single-point capabilities to complex task execution, and from occasional generation results to stable, scalable production capability.

In vertical domains, the value of Harness has been repeatedly verified.

In programming, there is Anthropic's Claude Code. Although it is powered by the Claude model underneath, its ability to read code, modify files, run tests, and continue fixing based on error reports all comes from the Harness layer outside the model. In website editing, Lovable lets users get a deployable website with a single sentence; after completing a $400 million Series C funding round in August this year, its valuation reached $13.3 billion. In design, Figma has a market value of approximately $11.54 billion (as of the close on November 2). Its real moat is no longer drawing features, but the fact that designers' asset libraries, version histories, and collaboration workflows are all accumulated on the platform, making migration costs extremely high.

In the AIGC field, Haima Cloud's RunningHub is a typical example of Harness in the field of image, audio, and video content production.

In the image and audio-video fields, the threshold for Harness is much higher. Because the demands of visual production are divergent, a single dialog box simply cannot complete delivery. It requires a longer toolchain and thicker context, from LoRA fine-tuning and complex scheduling pipelines to professional software such as Blender and Unreal, and further down to the underlying compute that integrates rendering and inference—none of these can be missing.

RunningHub, under Haima Cloud, is building exactly such a Harness for image, audio, and video production, with products such as Runow, the infinite canvas rhTV, and RHSTORY as its main carriers. The core of this Harness consists of three parts:

In Haima Cloud's view, the future will be an era in which Agents drive everything:

With sufficient context (Context is Enough) and a sufficient toolchain (Toolchain is Enough), an Agent can accomplish everything.

Following this logic, all interaction entry points on the platform will ultimately be upgraded into Agent entry points.

This is first reflected in RunningHub's dynamic model routing and intelligent orchestration. For the same requirement, which model to use, in what order to call them, and how much to spend used to require users to figure out on their own; now, these decisions are being handed over to the Agent. According to Haima Cloud, about 60% of model calls on RunningHub are already completed through intelligent dynamic routing: the Agent first identifies the user's intent, then combines contextual knowledge from industry know-how to decompose the task, assigning simple tasks to small models and complex tasks to large models. Different types of tasks, such as e-commerce images, cinematic visuals, and short comic-drama production, are also assigned to the most suitable models respectively.

In the new version of Runow, this capability has been built directly into the product. When "model preference" is set to "automatic," the Agent automatically selects Skills and models according to the task, so users do not need to first learn every model and workflow; in this way, quality is guaranteed and costs are kept under control—steps that should not use large models do not need to spend large-model money.

Of course, for an Agent to truly finish the work, two prerequisites are indispensable: first, sufficient context; second, a sufficient toolchain.

The toolchain is the necessary condition for giving the large model, the "brain," hands and feet. Different creation pipelines require different toolchains to be invoked, and among them the most complex is film-grade production.

Film-grade production has extremely high requirements for final output quality. Character assets, scene construction, and camera movement and scheduling all have strict standards at every stage, and it is almost impossible to complete production solely by conversing with a large model. To this end, Haima Cloud has connected multiple toolchains, including the canvas, LoRA training, ComfyUI workflows, and Blender, Figma, Premiere + cloud computer, going deep into the AIGC creation pipeline and building a film-grade production pipeline.

Among these, the Blender + cloud computer solution is particularly worth mentioning. Blender is essential industry software for 3D modeling and scene scheduling. Most "3D director stages" offered by AIGC platforms on the market are largely simplified toys aimed at beginners; once professional modeling is involved, users must repeatedly jump back and forth between the AI platform and Blender, breaking the production process.

RunningHub's approach relies on the heterogeneous container images and application-level unified scheduling capabilities that Haima Cloud has accumulated in its cloud computer business, directly deploying native professional software such as Blender—which is GPU-intensive and extremely latency-sensitive—on edge nodes, allowing the Agent to invoke it seamlessly within RunningHub's canvas.

Users only need to describe their requirements in natural language in the canvas, and the Harness (system orchestration) will dispatch the Agent to automatically open the Blender project in the cloud computer and execute the build; if fine adjustments are needed, creators can take over the same project at any time for manual fine-tuning.

The value of doing this is not only efficiency. On the one hand, the willingness of professional users to pay is on a different level from that of ordinary users, just as there is a difference between users of photo-editing apps and users of Photoshop; on the other hand, every operation by professional users in native Blender will be retained as expert trajectories on the platform, which in turn become nourishment for the evolution of the Agent's capabilities—and this is precisely what platforms that only offer simplified tools find very difficult to obtain.

The production of film-grade character assets is another typical scenario.

English text is being prepared.

In RunningHub, however, these steps can all be handed to an Agent through natural language. Using the film-grade material standards and reference images in the context, the Agent calls tools to automatically cut out images, selects suitable LoRAs or models to repair each local part, and finally calls tools to composite them. The director only needs to confirm at key checkpoints, and the cost is reduced by 90%.

From rhTV, which connects the film-grade production pipeline, to RHSTORY, which compresses the short-drama production cycle to "one person, one day," to the e-commerce Skill that precisely breaks down hit elements, Haimayun's Harness is enabling AI to truly grow "hands and feet" and reach deep into real production pipelines with high barriers to entry.

Whether this Harness has truly entered the production process—customers' continued investment is the best answer. After an e-commerce AI service provider connected to RunningHub at the end of last year, within one month it built 47 self-built applications, covering the entire production process from product images to short videos, and provided them to more than 100 merchant accounts under its umbrella; in the past two months, they called every day for 56 consecutive days, with cumulative calls exceeding 120,000 and monthly investment growing by 700%.

However, whether it is the Agent's intelligent orchestration or the connected invocation of the toolchain, both depend on a common prerequisite—sufficient context and data. In Haimayun's ecosystem, a sound Harness lets the large model grow hands and feet, while data is the blood that drives this engine to keep running.

Tools and processes can ultimately be imitated. The answer Dang Jinfeng gives for Haimayun's real moat is data:

A sound Harness is a necessary condition that lets the large model grow hands; data is another necessary condition, and together they constitute the sufficient and necessary conditions for the AI productivity system to operate continuously.

But the data Haimayun emphasizes is not generalized image libraries publicly scraped from the internet, but extremely valuable production-site data and expert decision trajectories obtained under compliant circumstances.

Precipitation of expert decision trajectories: When a senior creator completes a production task on RunningHub, which model they choose, which tool nodes they invoke, whether they first set up shots in the 3D director's console or build the process node by node on the canvas, or directly hand it to the Agent for one-click generation; how many rounds a video was revised, and whether each round adjusted lighting or expressions... these desensitized operation trajectories are fully recorded.

And these desensitized high-dimensional human expert interaction sequences are something that almost never appears in the training data of general large models. In vertical self-developed post-training, they are converted into reinforcement learning reward signals, forcing the model to fold and compress the originally cumbersome ten-step game into a single smooth native inference.

Assetization of industry know-how: The art style control of game art, the hit conversion formulas of e-commerce operators, and the camera language of film directors—these assets contributed with creator authorization are being structurally refined into reusable Skills and structured knowledge bases. As one of the earliest AIGC open-source communities in China, RunningHub has now accumulated millions of open-source workflows, still increasing at a rate of a thousand per day.

These data will ultimately flow into the knowledge base. The knowledge base is a structured database for AI invocation and a core component of the entire system. It is jointly built by the platform and ComfyUI and LoRA creators, covering industries such as architecture, gaming, film and television, and e-commerce.

At present, Haimayun's knowledge base is divided into two parts: one is product knowledge, namely RunningHub's rules and usage methods; the other is creation knowledge, namely prompts and production methods for images and videos.

Next, Haimayun will also build databases by industry, starting with e-commerce, film and television short dramas, and advertising. The material standards mastered by the Agent in the film and television case above come precisely from such knowledge accumulation.

Closed-loop self-evolution and high retention: Revisions and evaluations in real production tasks flow back as evaluation sets, feeding back into model post-training and Agent evolution.

On the one hand is the evolution of the model: production trajectories and expert experience obtained compliantly will be continuously used for post-training of the RH series models, making the models increasingly close to real production needs.

On the other hand is the evolution of the Agent: it will increasingly understand requirements and users, and require less and less rework. When a user says "make it a bit more refined" or "give it more texture," the Agent can select the right method and model, and even proactively recommend suitable materials, so users no longer need to study prompts; the more materials, brand assets, and historical projects a user accumulates, the closer the Agent's output will be to their style and specifications; and for common problems found in real conversations and task evaluations, once fixed, the proportion of similar tasks done right on the first try will also keep rising. Understanding ambiguous intent is precisely one of the most difficult and most valuable steps in professional production.

In addition, the ComfyUI workflows, trained LoRAs, and VI specifications that enterprises accumulate on the platform over time create extremely high switching costs. For example, after a brand designer hands the VI specifications to the Agent, every image subsequently generated will automatically comply with the brand specifications. Whether the platform can manage and reuse these assets well will directly determine whether customers stay. According to Haimayun, users with more accumulated self-built workflows have a retention rate close to 100%.

Among them, after an architectural design company adopted RunningHub, it built 7 applications and 14 workflows, including "structural repair," to address pain points in architectural renderings; its monthly investment grew more than 10-fold, and it treated the platform almost as a "second office." A mobile game development company, meanwhile, trained more than 80 dedicated LoRAs and built more than 160 self-built workflows on the platform, training its own game's art style into dedicated models covering tower defense scenes, chibi characters, game icons, and Chinese-style patterns, while also building production tools such as line-art refinement, motion transfer, multi-angle image generation, and image-to-3D, turning the entire game's art style into assets; usage grew nearly 70-fold within three months.

As Dang Jinfeng sees it: the reason consumer-grade AI does not make money is essentially that ordinary C-end users simply lack sustained and strong motivation to pay.

The internet competitive ecosystems in China and the United States are worlds apart; China's internet penetration depth, business model maturity, and evolution of user habits are far ahead of the rest of the world. Many C-end consumer models that rely on novelty, memes, or shallow packaging stacked up overseas (for example, overseas social and filter products such as Instax/Instinct that explode on short-term hype) simply cannot establish a sustainable commercial profit equation when they land in the extremely brutal competitive domestic market.

As vanity benchmark scores and shallow C-end applications gradually fade, what truly has high barriers and high customer renewal stickiness will inevitably be vertical AI productivity that delivers clear cost reduction and efficiency gains for B-end enterprises and professional creators.

Looking at Haimayun's overall strategic paradigm, its core advantage does not lie in single-point breakthroughs in a single dimension, but in building from the bottom up a highly self-consistent five-layer capability stack:

As AI commercialization enters deep waters, shallow applications that merely wrap APIs are destined to be passersby, and lab players that only pursue vanity benchmark scores with self-developed large models will also find it hard to escape the implementation dilemma. Only companies with full-stack five-layer capabilities from compute scheduling to application deployment, companies that build an AI productivity engine, can truly turn technology into sustainable commercial certainty and firmly secure a place at the core table of the final industry landscape.