BooleanMix's Four Rounds of Iteration: Using the TOLD Tech Stack to Solve the Challenge from Heterogeneous Computing Power to Token Delivery

In its report "Research Report on the Development of Intelligent Computing Infrastructure for Token Services (2026)," the China Academy of Information and Communications Technology (CAICT) offers a clear judgment: for the first time, inference computing power has surpassed training computing power to become the main source of demand, and Tokens are replacing FLOPS and the number of GPU cards to become the new yardstick for measuring the value of the AI industry. When reviewing emerging Token service providers, the report writes that Bolun Zhihui "focuses on heterogeneous intelligent computing platforms and gradient Token factory technical services, and its self-developed TOLD technology stack has been verified in real-world settings on a heterogeneous cluster of NVIDIA and domestic GPUs at the ten-thousand-card level." In CAICT's observation of the representative forces in the Token track, Bolun Zhihui enters this industry landscape alongside companies such as SiliconFlow, Qujing Technology, Qingcheng Jizhi, and Infinigence AI, which means that the proposition of "how heterogeneous computing power can be transformed into deliverable Tokens" is rising from an engineering issue to a public issue at the infrastructure level.

The difficulty of a Token factory does not lie in "whether there are GPUs." After racks are installed in many intelligent computing centers, they remain in a low-load state, and heterogeneous chips are difficult to coordinate with one another due to architectural differences. Between the chip and the stable output of Tokens lie multiple gaps: management and control, scheduling, adaptation, and operations and maintenance. Bolun Zhihui's BooleanMix platform is precisely the four rounds of iteration developed around this gap.

Four rounds of product iteration at Bolun Zhihui: an evolution path aimed at Token delivery

English text is being prepared.

The second launch pushed the focus toward the depth of adaptation between models and domestic computing power. The platform built an image and video generation service framework based on Diffusers, advanced inference engine adaptation for the Tianshu series and Hygon BR series chips, and introduced custom startup parameters for model services and a deployment verification step—test results are written into the Manifest, and the startup time for specific model and GPU combinations is also recorded and accumulated. On the operations side, model installation error logs, inference Pod logs, text-to-video monitoring, and secure download of video files were added in sync. From adaptation and configuration to verification and log monitoring, this round carried technical work through the entire process before and after model service operation.

The third launch was relatively focused, but it pointed to an easily overlooked delivery step: tenants can modify the Client-level background, copy, and Logo in the workbench, thereby connecting the underlying model services with the brand expression aimed at B-end users.

The fourth launch responded to the continuity challenge brought by continuous model updates. Inference Service added seamless upgrade capability, supporting non-stop updates after model optimization and cleanup of useless resources; the project ownership of inference framework images was standardized, and image repositories from the center to the edge are synchronized on demand, reducing unnecessary transmission and storage occupation. The new service management component is compatible with old versions of services, and existing services do not need forced upgrades and can be switched gradually. The new model launch process was also standardized, with adaptation, image preparation, service updates, and version compatibility organized into the same continuous delivery chain. The Benchmark testing platform, stress testing tools and Jenkins integration, and ARM multi-architecture CI/CD transformation also entered the R&D support lineup at this stage, accumulating verification tools for subsequent versions.

The TOLD technology stack: reversing the scheduling order from "use whatever you have" to "provide whatever is needed"

Behind the four iterations is a unified architectural main line: Bolun Zhihui's TOLD technology stack (Token-Oriented Large-scale Distributed Architecture). Traditional computing power scheduling is resource-oriented—first look at what GPUs are on hand, then figure out how to adapt and deploy; TOLD reverses the order, first understanding what kind of Tokens need to be delivered, then working backward to derive computing power, models, and scheduling strategies. Using logistics as an analogy: resource-oriented is like a company that only has trucks and uses trucks for delivery regardless of the attributes of the goods; TOLD is closer to an intelligent scheduling center that first clarifies the timeliness and attributes of the "goods," then decides what vehicles to use and what routes to take, splitting and coordinating when necessary. Tokens are the "goods" being scheduled.

A technical path being validated by the industry

From CAICT's report citing Bolun Zhihui to IDC China's continued attention to heterogeneous computing power management and Token metering, one signal is clear: infrastructure competition in the Token economy will not stop at "who has more cards," but will come down to "who can stably, efficiently, and measurably transform heterogeneous computing power into Token delivery." The four iterations of Bolun Zhihui's BooleanMix did not attempt to solve all problems at once, but converged layer by layer along the sequence of resource management, adaptation verification, tenant delivery, and continuous upgrading. This path itself is an engineering answer to the "Token factory" moving from concept to operable infrastructure.

Looking ahead to the next stage, Bolun Zhihui is opening two types of cooperation directions: first, it welcomes IDC operators, intelligent computing centers, and asset holders with high-quality domestic GPU or NVIDIA GPU assets to cooperate around the unified management and control of heterogeneous computing power and tiered Token output; second, it welcomes investors and industry partners interested in jointly building high-quality Token factories to advance the large-scale delivery from computing power assets to Token production capacity. Against the backdrop of Tokens gradually becoming the basic metering unit of AI services, open collaboration at the heterogeneous scheduling layer may be more worth betting on than single-point expansion of computing power.