Understanding the Base Model Landscape: A Quantitative Illusion and Layered Realities
This article clarifies why the number of true base foundation model companies is limited despite the proliferation of "large model" products. It introduces a layered framework (L0 base models, L1 industry adapters, L2 application integration) to differentiate underlying capabilities from application-level solutions, and provides practical guidance for enterprises evaluating and selecting models based on task requirements and engineering costs.
Since the rise of conversational generative AI, the term “large model” has become ubiquitous in technical discourse and business narratives. A surge of specialized concepts—such as financial large models, industrial large models, weather large models, and government large models—has followed. Terminological expansion is not inherently problematic. However, when all “models” are discussed in the same context, markets easily conflate industry-adapted models, application-packaged models, and foundational base models, leading to the mistaken impression that many companies are building base models. In reality, practitioners note that the number of startups genuinely developing base foundation models is very limited. This observation is not merely a numbers game but reflects the combined effects of model layering, engineering barriers, and commercial logic.
The Core of the Quantity Debate: Mistaking Model Applications for Base Models
A base foundation model typically refers to a model pretrained on large-scale data to acquire generalizable representations, capable of transferring across multiple tasks and domains. Its essential characteristic is not simply a large parameter count or the inclusion of “large model” in its name, but rather strong generality, transferability, and continuous iteration capability. In contrast, many industry-specific large models are built by continued training, fine-tuning, knowledge enhancement, or workflow orchestration atop a base model, targeting specific scenarios.
Therefore, any discussion of “how many companies build base models” must first define the statistical scope:
- Base model providers: Responsible for large-scale pretraining, alignment, evaluation, inference optimization, and long-term version iteration.
- Industry adapters: Perform domain-specific data training, prompt engineering, retrieval augmentation, tool invocation, and delivery integration on top of a base model.
- Application integrators: Embed model capabilities into business processes, creating intelligent customer service, document processing, data analysis, code assistance, and other applications.
If the latter two categories are also counted as “large model companies,” the number naturally inflates. But the teams that truly possess the capability for sustained L0 base model R&D are typically a handful. This explains why “hundreds of models competing” and “limited base model supply” can coexist: the former describes the flourishing application and adaptation layer, while the latter describes the concentration of underlying general-purpose capabilities.
L0 and L1 Layering: Why Base Model Companies Are “Countable on One Hand”
Using L0 and L1 to differentiate clarifies the market structure. The L0 layer represents base models, emphasizing general pretraining ability, cross-task transfer, scalable inference, and long-term iteration. The L1 layer refers to industry or scenario models, usually built by continued training or adaptation on an L0 model, focusing on domain knowledge, task accuracy, compliance, and delivery efficiency. A further L2 layer (application and integration) productizes and operationalizes model capabilities.
| Layer | Primary Goal | Capability Source | Main Costs | Evaluation Focus |
|---|---|---|---|---|
| L0 Base Model | General understanding, generation, and reasoning | Large-scale pretraining, alignment, evaluation systems | Compute, data governance, training stability, talent | General capability, transferability, inference cost, iteration cadence |
| L1 Industry Model | Domain-specific task adaptation | Continued training, fine-tuning, knowledge enhancement on L0 | Domain data, annotation, evaluation, compliance | Task metrics, domain coverage, deliverability |
| L2 Application Integration | Business process deployment | Model invocation, tool orchestration, retrieval augmentation, workflows | System engineering, delivery, operations, security | User experience, stability, integration cost, ROI |
The high barrier of the L0 layer drives supply concentration. Large-scale pretraining demands sustained compute investment, data cleaning and governance, training stability engineering, evaluation systems, and safety alignment. Post-release, the model requires continuous iteration, inference cost reduction, and developer ecosystem maintenance. These are not one-time achievements but ongoing engineering efforts. Funding rounds ranging from hundreds of millions to billions of RMB, with a minority of teams reaching decabillion valuations, indicate that capital concentrates on a few base model teams. Yet concentrated funding does not automatically equal commercial maturity; the real metrics are unit inference cost, customer retention, delivery capability, and sustainable revenue.
The Real Problem of Terminology Inflation: Difficult User Evaluation and Harder Customer Acquisition
When “large model” is broadly used to name industry solutions, buyers first face blurred capability boundaries. A “government large model” might merely be a knowledge-base retrieval and QA orchestration on a general model; an “industrial large model” may primarily include equipment knowledge QA and report generation. These solutions have value, but they are not the same as L0 base models. Terminology inflation increases communication costs: users struggle to determine whether a vendor supplies an underlying model, an industry adapter, or an application integration. Meanwhile, upstream base model teams struggle to be seen and accurately assessed amid the noise.
To address this, the industry should gradually adopt clearer communication practices:
- Define the layer: Specify whether a solution is a base model, an industry adapter, or an application integration, avoiding a single term that covers different capabilities.
- Publish evaluation criteria: Show not only general benchmarks but also scenario-specific tasks, data conditions, inference costs, and failure cases.
- Provide model cards and data cards: Describe training data scope, applicable scenarios, limitations, safety policies, and version changes.
- Distinguish demo capability from delivery capability: Demo effects do not equal engineering stability; pay attention to concurrency, latency, observability, and operational overhead.
From Visibility to Customers: The Engineering Path for Base Model Companies
To be recognized in a confusing market, base model teams must convert capabilities into verifiable engineering interfaces and delivery paths, rather than relying on concept announcements alone. Common practices include providing APIs, SDKs, private or hybrid deployment options, model version management, transparent usage and cost metrics, and developer-facing documentation and examples. For enterprise customers, what matters most is not the “size of the model name” but whether it can match target metrics on specific tasks within an acceptable cost.
On the engineering side, cloud platforms—including Baidu AI Cloud—typically support model evaluation and delivery pipelines with general components such as container platforms, object storage, logging services, monitoring and alerting, and message queues. These generic capabilities do not determine model strength but significantly affect iteration efficiency and operational quality. For example, evaluation tasks can run in containers, datasets can be stored in object storage, inference services can be observed through logs and alerts for latency and error rates, and asynchronous tasks can be smoothed by message queues.
def choose_model_layer(task):if task.scope == "cross_domain" and task.needs_general_reasoning:return "L0_base_model"if task.domain_specific and task.has_labeled_data:return "L1_industry_adapter"return "L2_application_orchestration"def evaluate_candidate(model, eval_set):metrics = {}for case in eval_set:output = model.generate(case.input)metrics[case.name] = score(output, case.expected)return metrics
The pseudocode above illustrates a basic selection logic: for cross-domain general tasks, prioritize evaluating L0 base models; for domain-specific tasks with labeled data, evaluate L1 adapters; for complex business workflows, rely on L2 application orchestration. The key is that evaluation sets must come from real tasks, not just generic samples.
Multimodal Evolution and Iteration Pace: Capability Expansion Is Not Narrative Expansion
The release of multimodal base models expands capability stacks from text to images, documents, audio, and video, but also introduces more complex evaluation and engineering challenges: cross-modal alignment, data cleaning, inference costs, safety filtering, and version compatibility require systematic handling. When evaluating multimodal capabilities, enterprises should decompose tasks into measurable subtasks—such as document structure understanding, chart QA, image description consistency, and cross-modal retrieval accuracy—rather than relying solely on demo effects.
Enterprise Selection Framework: Start from Tasks, Not Concepts
Given the noise of “hundreds of models,” enterprises can adopt the following selection framework:
- Define tasks and boundaries: Separate general reasoning, domain QA, document processing, code assistance, multimodal understanding, and other tasks.
- Layer selection: First determine whether an L0 base model is needed, or whether L1 industry adaptation or L2 application orchestration is sufficient.
- Build a private evaluation set: Use real business samples covering normal, edge, and failure cases.
- Calculate total cost: Include inference cost, adaptation cost, operational cost, compliance cost, and migration cost.
- Design an exit mechanism: Avoid over-coupling to a single model version; preserve the ability to migrate prompts, adapters, and data assets.
- Continuously monitor: Track latency, error rate, hallucination rate, refusal rate, and user feedback after deployment.
Conclusion
The limited number of base foundation model startups does not reflect a lack of market demand but rather the high L0 R&D barrier, high iteration costs, and concentration of capital and talent. Meanwhile, L1 industry models and L2 application integrations will continue to thrive, forming a layered ecosystem of “few but strong base models, many but specialized industry adapters, and broad and deep application integrations.” For developers and corporate decision-makers, the essential move is not to chase every “large model” concept but to identify its layer, validate real task performance, account for engineering costs, and complete selection and deployment under a clear layered logic.
