Where to find models and what's out there: one combined overview
The landscape of artificial intelligence models has become highly fragmented in recent years. Where organizations previously relied mainly on a small number of large API providers, today's ecosystem consists of a layered structure of open-weight repositories, managed cloud platforms, specialized hosting services, and local execution environments. Making a well-informed choice for a specific application requires a clear overview of where these models are made available and the categories into which they fall.
This overview brings together the most important distribution channels and model types. It acts as a neutral compass for software developers, data architects, and IT decision-makers who want to select infrastructure without becoming dependent on marketing terms. To understand where model distribution fits within the broader classification of software solutions, you can look at the AI ecosystem mapped out, which clearly organizes all layers from infrastructure to application.
Classification of the model landscape: four distribution forms
AI models are not offered in one uniform way. In practice, we distinguish four primary distribution forms, each with its own architectural characteristics, cost structure, and management requirements. Understanding these forms prevents incorrect assumptions about privacy, scalability, and intellectual property.
The first category concerns hosted proprietary APIs. Here, the model architecture and the weight set remain entirely in the provider's possession. Users send input via an encrypted network connection and receive the generated output back. The provider bears responsibility for hardware, optimization, and scalability, but the customer has no direct control over the underlying code or model changes.
The second form is that of the open-weight model hubs. On these platforms, research institutes and tech companies publish the exact weights of trained models. Developers can download these weights and host them themselves. However, a sharp distinction must be made here between 'open weight' and 'open source': for many open-weight models, the dataset and the specific training code are not released, meaning full reproducibility is lacking.
The third category is cloud-managed infrastructure (Model-as-a-Service). Large cloud providers offer curated catalogs of both proprietary and open-weight models. The models run within the customer's own cloud environment, so data never leaves the trusted perimeter. This combines the flexibility of open models with the operational convenience of a managed API.
The fourth form is on-device and local distribution. Here, compressed or quantized versions of models are distributed via specific runtime packages to run directly on end-user hardware, such as laptops, smartphones, or edge servers. Ruling out external network connections this way offers maximum privacy but places strict limits on model size.
Model hubs and open-weight repositories
For organizations that want full autonomy over their AI pipeline, open-weight repositories are the most important source of models. These platforms act as the library and marketplace of the AI ecosystem. For a deeper look at repositories that specialize in hosting weights and training sets, you can turn to the article on where to find models and datasets.
The best-known and largest platform in this category is Hugging Face Hub. This platform works as a central registry for hundreds of thousands of models, datasets, and interactive demos. Hugging Face offers standardized libraries (such as transformers) that let developers load a model with just a few lines of code. Besides Hugging Face, platforms such as Kaggle, GitHub and vendor-specific portals (such as the repositories of Meta, Mistral AI, or Alibaba Cloud) play an important role in the initial release of new model generations.
.bin or .pkl formats) can execute arbitrary code when loaded. Preferably use safe file formats such as .safetensors to rule out security risks.
The main advantage of model hubs is transparency. Users can precisely analyze how large a model is, which license terms apply, and which community evaluations are available. The weakness, however, lies in the variables: a model's performance from a hub depends heavily on the chosen hardware, the quality compression used (quantization), and the runtime software.
Proprietary API providers and closed ecosystems
When maximum compute power, complex reasoning abilities, or extensive multimodal capabilities are required without your own hardware investments, closed API providers are the default choice. These providers develop very large models whose exact size and architecture are often trade secrets.
Pioneers in this space are organizations such as OpenAI, Anthropic, and Google (via Vertex AI and AI Studio). Their model portfolios are divided into different performance tiers: from very fast, cheap models for routine tasks to large reasoning models that perform multi-step analyses. Besides these established players, companies such as Cohere and AI21 Labs offer models specifically optimized for business search applications and document processing.
The advantages of closed API services are a low entry barrier and constant performance improvements without any maintenance of your own. There are, however, clear downsides:
- Provider Dependency (Vendor Lock-in): Models can be updated or withdrawn by the provider, which affects the reproducibility of results.
- Data Processing: Input and output data leave the organization and are processed on external servers, which requires additional legal safeguards.
- Cost Predictability: Billing is based on the number of tokens (text segments) processed, which can lead to budget overruns in the event of unexpectedly high volumes.
Model families classified by functionality and modality
When selecting a model, the way information is input and processed is of primary importance. Models are no longer limited to purely textual processing; there is now a wide differentiation in modalities.
| Model Category | Input / Output | Typical Applications | Known Category Examples |
|---|---|---|---|
A specific subcategory that is quickly gaining ground is the embedding models. These models don't generate readable text but convert input into numerical vectors that represent the semantic meaning of a text or image. They form the crucial link in RAG systems (Retrieval-Augmented Generation), where external knowledge sources are connected to generative models.
Open-weight model families and their license structures
Within the domain of freely accessible weights, a number of defining model families have emerged. Choosing a specific family has direct consequences for legal permissibility within commercial products. For a detailed analysis of the best-known open weights and their license structures, you can consult the overview of open source LLMs at a glance is worth consulting.
Important players in this category include the Llama series from Meta, the Mistral and Mixtral models from Mistral AI, the Qwen family from Alibaba Cloud, and the Gemma models from Google. Although these models are often referred to as "open source", the issuing organizations apply widely differing terms:
- Permissive Open Source Licenses (e.g., Apache 2.0, MIT): Grant full freedom for commercial use, modification, and redistribution without limits on the number of users. This applies, for example, to many Mistral AI models and various Qwen variants.
- Restricted Commercial Licenses: Grant the right to commercial use but impose conditions. For example, the license for Meta's Llama series requires an additional license once the application exceeds a specific number of monthly active users, and it prohibits using the model output to train other language models.
- Non-Commercial Research Licenses: Strictly limit use to academic or personal evaluation. Commercially deploying processed data or running the API for customers is explicitly prohibited under these licenses.
Local execution and runtime infrastructure
Owning model weights is only the first step; running a model locally or on your own servers requires specific runtime software. This software translates the model's mathematical operations into efficient instructions for the available compute hardware (such as GPUs, NPUs, or CPUs).
When you want to run models on your own hardware without depending on external cloud providers, the overview of tools for running LLMs locally offers a practical starting point. Popular solutions in this category include:
- Ollama: A widely used runtime that simplifies installing and managing quantized models through a clear command-line interface and a local API.
- LM Studio: A graphical application for desktop environments that lets users quickly test models, adjust parameters, and start a local OpenAI-compatible server.
- vLLM and TGI (Text Generation Inference): High-performance, specialized engines designed for production-grade server environments. They use advanced memory techniques (such as PagedAttention) to efficiently handle hundreds of concurrent requests.
- llama.cpp: The underlying C/C++ engine that makes it possible to run models with minimal hardware requirements and efficient quantization on a wide range of processors.
If you're looking for practical guidance on retrieving, storing, and managing weight files on local infrastructure, consult the guide on downloading and managing local models.
Comparison of distribution channels: pros and cons
To determine which distribution channel best fits an organization's objectives, the operational characteristics need to be systematically compared. The table below compares the three main distribution routes across key evaluation criteria.
| Criterion | Proprietary APIs | Cloud-Managed Open Weights | Own / Local Hosting |
|---|---|---|---|
Selection criteria, governance, and data residency
Choosing the right model channel is rarely purely technical; legal and operational boundary conditions play a decisive role. Organizations operating within the European Union must specifically take into account the provisions of the General Data Protection Regulation (GDPR) and the European AI legislation (EU AI Act).
When using proprietary API services, it's important to carefully check where data processing takes place. Many US providers offer processing in the United States by default, which can lead to compliance issues without additional legal agreements (such as the Data Privacy Framework or standard contractual clauses). Managed cloud platforms in European data centers offer better guarantees on data residency in this regard.
A second important selection criterion is measuring the effect of model optimization. Many open-weight models are reduced in size (quantized) to fit on cheaper hardware. This, however, has an impact on model accuracy. For organizations that want to understand how much precision is lost when weights are compressed for local use, the benchmark page on the quality impact of quantization analyzes the exact performance differences.
If you're unsure which type of solution fits your specific work environment and budget, the AI Tool Picker helps you select the right channel and corresponding architecture step by step.
Step-by-step plan for model selection and implementation
To arrive at a responsible model choice within an organization, it's advisable to follow a structured selection procedure. This method prevents choices from being based solely on brand recognition or temporary hype.
- Classify data sensitivity: Determine whether the information to be processed is public, internal, confidential, or strictly secret (e.g., medical data or trade secrets). This step immediately rules out certain distribution channels.
- Analyze the required functionality: Determine the necessary context length, modality (text only or also images), and the desired response speed.
- Start with a proprietary API as a benchmark: Build an initial prototype with a high-quality commercial API to test whether the idea is feasible in substance and what the maximum quality is.
- Evaluate open-weight alternatives: Test whether a smaller, hosted open-weight model (such as a quantized Llama or Qwen variant) achieves a comparable quality level at lower operational cost or with better privacy safeguards.
- Formalize MLOps and governance: Document how model updates are tested, how output quality is periodically monitored, and which fallback options apply if a provider goes down.
Conclusion and maintenance advice
Today's AI model landscape offers a suitable distribution channel for every business scenario. While hosted APIs offer the fastest path to innovation, open-weight repositories and local runtimes lay the foundation for privacy, autonomy, and long-term cost control. A successful AI strategy rarely relies on a single model, but instead uses a hybrid architecture in which different model types are deployed side by side based on task complexity and data sensitivity.
Since developments in the AI sector move very quickly and model families are refreshed monthly, it's essential to periodically review model choices and abstract dependencies in the software code as much as possible through universal API adapters.
