China's LLM "Hundred-Model War": Landscape Evolution and the Path to Industry Divergence

    This article examines the rapid rise of large language models (LLMs) in China, tracing the shift from a prolonged "AI winter" to the concentrated emergence of more than a hundred models in under a year. It breaks down the market structure across parameter scale, domain distribution, and application form, and explains why parameter count alone is an unreliable measure of model strength. It then contrasts two parallel engineering philosophies—general-purpose foundation models versus vertical, industry-specific models—and offers a neutral view of the differences between US and Chinese development paths. Finally, it maps the vertical differentiation that drives real industrial adoption and walks through the full engineering pipeline from data governance and training to retrieval-augmented generation (RAG) and inference deployment, including cost and observability considerations.

    1. From the “AI Winter” to a Concentrated Wave of Large Models

    For a long stretch of time, the field of artificial intelligence struggled to translate technical progress into commercial value. Many technical directions could not sustain viable revenue, the funding environment tightened, and a large number of projects gradually faded from competition. Capital and talent became concentrated in just a few hot segments.

    What truly changed this landscape was the breakthrough in generative large model capabilities. Notably, the core of this shift was not a “disruptive leap” in the underlying model technology itself, but rather the re-ignition of the commercial environment. The general-purpose capabilities demonstrated by large models allowed the market, for the first time, to clearly see the potential for large-scale application—giving rise to an industry landscape in which new growth flourished alongside the old.

    The significance of this shift lies in AI moving from “technically feasible” to “commercially tangible.” When developers can invoke powerful capabilities through natural-language interaction and quickly package them into applications for specific scenarios, the expected return on technical investment across the entire industry is reshaped significantly.

    2. The Scale and Structure of the “Hundred-Model War”

    According to incomplete public industry statistics, the number of large models released in China surpassed the hundred-model mark in less than a year. The structure can be clearly divided into several layers:

    • By parameter scale: Models with 10 billion (10B) or more parameters already account for a considerable share, indicating that the industry as a whole has entered a “large-parameter” competitive range.
    • By domain distribution: General-domain large models make up only a limited proportion, coming mainly from large internet platforms with accumulated compute and data resources, as well as research institutions.
    • By application form: Most of the remaining models focus on vertical industries, adapting to specific scenarios such as finance, healthcare, law, manufacturing, and education.

    This “pyramid-shaped” structure reveals a key judgment: competition in large models has moved from “whether one exists” to “how deep it goes.” The industry is shifting from horizontally covering general capabilities to vertically going deep into specific scenarios.

    How to Rationally Assess Parameter Scale

    When discussing “whose model is stronger,” parameter count is often treated as the core metric—but this understanding needs more rigorous technical grounding. A model’s final capability depends not on parameter count alone, but on the matching relationship among three factors: parameter scale, training data volume, and compute investment. The widely accepted Scaling Law shows that stable performance gains appear only when all three grow in concert. Piling up parameters while lacking high-quality data support often leads to “wasted compute” or even performance saturation.

    A more reasonable set of dimensions for evaluating a large model includes:

    1. The scale, quality, and diversity of the training corpus;
    2. The completeness of the training strategy (pre-training, instruction fine-tuning, reinforcement learning from human feedback, and so on);
    3. The cost, latency, and stability at the inference stage;
    4. Actual performance on specific tasks, rather than benchmark scores alone.

    3. General-Purpose Models and Vertical Models: Two Parallel Routes

    The divergence in approaches to large models is essentially a difference between two engineering philosophies: a “general-purpose capability platform” versus an “industry-specific model.”

    General-purpose large models pursue breadth of capability. The goal is to build a foundation with broad world knowledge and reasoning ability that can cover diverse tasks such as question answering, writing, coding, translation, and logical reasoning. The challenge lies in controlling hallucinations and improving the reliability of instruction following while maintaining generality. These models are typically built by teams with large-scale compute, data governance, and distributed training experience.

    Vertical large models center on industry depth. Through continued pre-training on domain data and task fine-tuning, they “reshape” a general foundation into an expert for a specific scenario. For example, a model for legal scenarios needs to strengthen statute retrieval and case-law reasoning, while a model for healthcare scenarios places greater emphasis on terminology accuracy and risk control. The value of a vertical model lies not in larger parameters, but in usability, controllability, and compliance on specific tasks.

    A pragmatic technical route is often a combined strategy—“starting from a general-purpose foundation and landing on vertical fine-tuning”—rather than an either/or confrontation.

    4. Differences Between US and Chinese Development Paths (A Neutral View)

    Discussion of the gap between US and Chinese large models never stops, yet views rarely converge—because the evaluation dimensions differ. Some measure by parameters and benchmark scores, others by engineering efficiency and application density. A more objective perspective is to acknowledge the differences in development stage and areas of focus between the two, rather than simply declaring one “ahead” or “behind.”

    • In terms of building foundational capabilities, compute scale, the reserve of top-tier talent, and the depth of the data ecosystem are the core variables that shape the long-term ceiling.
    • In terms of application deployment efficiency, the ability to quickly turn model capabilities into deliverable products and industry solutions is another key competitive factor.
    • In terms of engineering and cost optimization, inference cost, deployment flexibility, and the speed of scenario adaptation determine whether a model can truly scale.

    Teams in different regions tend to emphasize different dimensions depending on their resource endowments and market demands. Understanding competition as “parallel exploration of diverse routes” is closer to the technical reality than the linear-ranking “catch-up” narrative.

    5. Vertical Differentiation: The Real Watershed for Industrial Adoption

    From a competitive standpoint, the teams that can weather the cycle are those that deeply embed technical capabilities into industry workflows. The core logic of vertical differentiation is as follows:

    1. Data moats: Industry-proprietary data forms a defensive barrier that general models struggle to replicate.
    2. Closed-loop scenarios: Model capabilities produce quantifiable value only when embedded in concrete business processes (such as contract review, fault diagnosis, or intelligent customer service).
    3. Compliance and controllability: For heavily regulated industries, output stability, traceability, and data security become entry barriers.

    This also explains why vertical models rapidly outnumber general-purpose ones—the real demand of the industry is far more pressing than “building a bigger model.”

    6. Engineering Practice: The Full Pipeline from Training to Inference

    Whether general-purpose or vertical, deploying a large model follows a similar engineering pipeline. The illustrative technical architecture below helps developers build a complete picture:

    1. [Data Collection & Cleaning] → [Pre-training / Continued Pre-training] → [Instruction Fine-tuning] → [RLHF Alignment]
    2. ↓ ↓
    3. [Vectorization & Knowledge Base] ←──────────── [Retrieval-Augmented Generation (RAG)] ←──── [Inference Serving]
    4. ↓
    5. [Monitoring & Alerts · Log Tracing · Cost Accounting]

    Key engineering stages:

    • Data governance: Cleaning, deduplication, and anonymization are the first line of quality assurance, directly determining the ceiling of subsequent capabilities.
    • Training and alignment: Pre-training provides foundational capabilities, instruction fine-tuning improves task adherence, and feedback-based alignment optimizes output safety.
    • Retrieval-Augmented Generation (RAG): Supplementing a model’s limited up-to-date knowledge with an external knowledge base is a mainstream approach for mitigating hallucinations and improving controllability.
    • Inference deployment: Container platforms enable elastic scaling, while monitoring, alerting, and logging services ensure the stability and observability of online services.
    • Cost management: Inference cost often dominates operating expenses; model quantization, batching, and caching are common optimization techniques.

    At the infrastructure level, the common industry practice is to use object storage to hold training data, message queues to decouple task flows, container platforms to host inference services, and logging and monitoring services to achieve end-to-end observability. These are standard cloud infrastructure categories that teams can combine flexibly according to their own scale.

    7. Conclusion

    China’s “Hundred-Model War” appears on the surface to be a competition over quantity and parameter scale, but in essence it is a dual contest between the ability to build general capabilities and the ability to deliver industry adoption. Rationally viewing the differences between US and Chinese routes, focusing on deep adaptation to vertical scenarios, and investing in foundational capabilities such as data, engineering, and cost are the keys to weathering this competitive cycle. For developers and enterprise users, rather than fixating on single-point parameter rankings, it is better to establish a technology-selection framework that is “oriented toward scenario value and supported by the engineering pipeline”—this is where large models truly create value.