Chinese Open-Source Models Go Enterprise: From High Usage to High Value

    Enterprise adoption of large models is moving from API experimentation into core workflows, but rising token consumption is not translating into proportional revenue for model providers. This article examines why Chinese open-source models have gained usage share, how the real cost structure extends well beyond token prices, and what providers must build—model routing, retrieval augmentation, evaluation, and deployment support—to move from API supply toward enterprise value. It also outlines the main commercialization challenges and a practical evaluation framework for enterprises.

    I. Market Signal: Usage Is Growing, but Revenue Share Is Falling

    In recent years, the way enterprises use large models has been changing. In the early stage, enterprises often used a small number of API calls to validate capabilities such as text generation, question answering, and code assistance. As inference costs decline, context capabilities improve, and the deployment barrier for open-source models falls, large models are beginning to enter core processes such as customer service, R&D, data analysis, knowledge management, and business automation.

    Industry statistics show that global enterprise AI spending is still growing, and the share of open-source models in enterprise token consumption continues to rise. Notably, higher model call volume does not mean that model providers can obtain proportional revenue. One set of representative industry data indicates that open-source models’ share of enterprise token consumption rose from about 10% to nearly 20%, while their corresponding share of spending fell from about 6% to about 4%.

    This shows a clear usage–revenue gap in the market:

    • Enterprises are making more calls, and outputs per task are becoming longer.
    • Models are embedded in more business systems, increasing call frequency.
    • Per-token prices continue to fall, approaching infrastructure costs in some scenarios.
    • Enterprises are using multi-model routing, caching, and localized deployment to reduce overall spending.

    As a result, business models that rely solely on call-volume billing are facing greater growth pressure.

    II. Why Chinese Models Achieve Higher Usage

    Chinese models account for a relatively high share of overseas enterprises’ use of open-source models. This is not caused by a single factor but is the combined result of technology, cost, and delivery approach.

    1. Inference Costs Are More Compatible with Scale

    What enterprises really care about is not the number of model parameters but the comprehensive cost per task. For scenarios such as classification, summarization, information extraction, and internal question answering, medium-sized models can often already meet requirements. If a model performs reliably in inference speed, memory footprint, and concurrency, enterprises have an incentive to deploy it on lower-cost computing resources.

    When a model supports quantization, batch processing, and continuous inference, enterprises can further improve hardware utilization. For business systems that generate a large number of requests every day, even a small change in per-token cost can create a significant difference in monthly bills.

    2. Open Weights Reduce Vendor Lock-In Risk

    Enterprises are usually unwilling to bind core business processes completely to a single provider. Open-weight models allow enterprises to conduct private deployment, capability evaluation, and secondary integration under license terms, making them more suitable for organizations with high requirements for data security, compliance auditing, and system controllability.

    This does not mean that enterprises can simply download a model and put it into production. Real deployment also involves engineering work such as inference frameworks, model quantization, service orchestration, access control, log retention, and fault recovery. However, the open model at least gives enterprises more room for choice.

    3. Multilingual and Complex Business Scenarios Drive Adoption

    Cross-regional enterprises often need to process Chinese, English, and other language content at the same time, and they also need to adapt to business processes and document formats in different countries. Models that can balance multilingual understanding, structured output, and long-text processing are more likely to enter enterprise internal systems.

    However, general model capability does not directly equal business usability. Enterprises still need domain data evaluation, prompt constraints, retrieval augmentation, and human review to ensure that outputs meet actual requirements.

    III. The Cost Structure Behind Usage Growth

    Tokens are not the only cost, but they are the easiest metric to count and compare. After enterprises adopt large models, actual spending usually consists of the following parts:

    1. Model inference cost: including computing resources, memory, network transmission, and high-availability deployment costs.
    2. Data processing cost: including document parsing, chunking, vectorization, index updates, and data cleaning.
    3. System engineering cost: including gateways, authentication, rate limiting, retries, caching, monitoring, and alerting.
    4. Security and compliance cost: including sensitive information detection, access auditing, data isolation, and permission management.
    5. Human operations cost: including evaluation set maintenance, prompt iteration, exception handling, and business review.

    If enterprises compare only the unit price of a model, they often ignore the latter four costs. A seemingly cheap model may instead increase total cost of ownership because of unstable output, higher retry rates, or the need for extensive human review.

    Therefore, a more reasonable metric is “cost per effective task,” not “price per token.” For example, in customer service scenarios, enterprises can calculate the comprehensive cost of completing one accurate answer. In code assistance scenarios, they can calculate the call cost required to generate a code segment that passes tests.

    IV. From API Provider to Enterprise Solution

    As model calls become increasingly standardized, differences at the API level will shrink. To capture higher value, model providers need to move beyond “providing an inference interface” and toward “helping enterprises complete tasks.”

    1. Offer Model Routing

    Different tasks have different requirements for models. Simple classification, keyword extraction, and format conversion can be handled by low-cost models. Complex reasoning, long-document analysis, and high-risk decisions require stronger models.

    Enterprises can set routing policies in a unified gateway to automatically select models based on task type, context length, latency requirements, and budget. A basic routing logic can be abstracted as follows:

    1. Receive request
    2. ├─ Determine task type
    3. ├─ Check data sensitivity level
    4. ├─ Evaluate context length and latency requirements
    5. ├─ Select appropriate model or deployment location
    6. ├─ Execute caching, retries, and rate limiting
    7. └─ Record quality, cost, and latency metrics

    This architecture can prevent all requests from using high-cost models and can also improve system resilience when a single model fails.

    2. Make Retrieval Augmentation a Standard Capability

    Enterprise knowledge is usually scattered across documents, databases, tickets, and business systems. The model itself does not inherently possess this real-time information, so enterprise applications generally need to introduce retrieval-augmented generation workflows.

    A typical workflow includes:

    • Cleaning source materials and tagging permissions.
    • Splitting documents by semantics and structure.
    • Building vector indexes and keyword indexes.
    • Retrieving relevant content based on user questions.
    • Ranking and deduplicating retrieved results.
    • Providing permission-checked content to the model for answer generation.
    • Returning citations and recording audit information.

    The key to this workflow is not only whether content can be retrieved, but also whether permissions are correct, citations are complete, and the model avoids speculation beyond the source materials.

    3. Build a Continuous Evaluation System

    Model performance cannot rely only on human experience. Enterprises should build evaluation sets covering accuracy, stability, security, latency, and cost, and automatically run regression tests after model upgrades, prompt changes, or knowledge base updates.

    Evaluation can be divided into three layers:

    • Offline evaluation: using a fixed question set to compare different models and parameter configurations.
    • Canary evaluation: routing only some real requests to a new version.
    • Online monitoring: continuously observing failure rates, human transfer rates, user feedback, and cost per task.

    Only by linking model quality to business metrics can enterprises determine whether an upgrade brings real benefits or merely improves certain laboratory metrics.

    V. Three Challenges in Commercializing Open-Source Models

    1. Price Competition May Compress Margins

    When multiple models are close in capability, enterprises can easily reduce costs by comparing prices and switching providers. If model providers rely only on low prices to attract traffic, they may quickly gain call volume but struggle to establish stable profit.

    A more sustainable approach is to provide value-added capabilities around industry workflows, such as data governance, deployment management, evaluation services, permission systems, and operations tools. However, these capabilities require long-term engineering investment and cannot be solved simply by expanding model scale.

    2. Overseas Deployment Must Meet Different Compliance Requirements

    Enterprise data differs by region, industry, and sensitivity level. When model providers enter overseas markets, they need to clarify data storage locations, processing boundaries, log retention methods, and cross-border transmission mechanisms.

    For highly sensitive data, enterprises may require local deployment or dedicated isolated environments. For ordinary text tasks, they may accept remote calls. As a result, model services need to support multiple forms, including public cloud, dedicated environments, and localized deployment.

    3. Model Capability Is Not Business Outcome

    A model’s performance on public tests cannot fully represent its value in enterprise workflows. Enterprises focus more on task completion rate, cost of errors, human intervention ratio, and system stability.

    For example, a model that answers fluently but makes factual errors may increase the workload of customer service and review staff. A model with strong reasoning capability but excessively high response latency may also fail to fit real-time business needs. Therefore, commercialization must be designed around real workflows, not only around competition on model leaderboards.

    VI. How Enterprise Users Should Evaluate Model Solutions

    When choosing open-source models or related services, enterprises can establish a quantifiable evaluation framework:

    1. Task fit: whether the model is suitable for the target business, rather than looking only at general test scores.
    2. Deployment flexibility: whether it supports different approaches such as local deployment, dedicated environments, and elastic computing.
    3. Cost controllability: whether comprehensive costs can be reduced through caching, batch processing, quantization, and routing.
    4. Stability: whether latency, error rates, and fault recovery under high concurrency meet requirements.
    5. Security: whether it supports permission control, sensitive information handling, and complete auditing.
    6. Portability: whether applications can switch smoothly between different models.
    7. Evaluation capability: whether quality changes can be tracked continuously rather than accepted only once.

    In architecture design, enterprises should encapsulate model calls in a unified service layer and avoid having business code directly depend on a specific model API. This makes model replacement easier and also centralizes authentication, rate limiting, logging, and cost accounting.

    VII. Future Competition: From “Model Share” to “Task Value”

    The relatively high call volume of Chinese models in the global enterprise market shows that their technical supply has entered real business processes. But call volume is only the result of market penetration, not the endpoint of commercial success.

    Competition in the next stage will focus on three areas:

    • Who can complete enterprise tasks stably at lower cost.
    • Who can truly integrate models with data, processes, and permission systems.
    • Who can turn one-time calls into long-term, measurable business value.

    For model providers, what needs to be built in the future is not only an inference API but also model routing, knowledge base management, evaluation platforms, deployment tools, and operations systems. For enterprises, the focus should also not stop at comparing unit prices; they should build a complete loop from data governance to outcome acceptance.

    As large models gradually become enterprise infrastructure, market evaluation standards will inevitably shift from “who is called more” to “who can continuously create effective output.” This is also a key step for open-source models to move from traffic competition to value competition.