A model supplies a capability. An application makes it usable.
An AI model is trained software that produces an output from an input. A text model can draft an answer from a prompt; a speech-recognition model turns audio into text; an image model may generate a picture from a description. The model does not automatically include your files, a convenient interface or permission to act on your accounts.
An application connects those capabilities to a task. ChatGPT, Claude and Gemini provide interfaces, account features and tools around models. A document application can add file import and retrieval; an image editor can add layers and selections. The application and the model it uses are different choices, and a product may change models or offer several.
An agent uses a model to choose steps and call tools while pursuing a task. A coding agent might read a file, edit it, run a test and inspect the result. An ongoing personal agent may work across connected apps. Tool access makes these systems more useful, but it also makes permissions, review and stopping conditions part of the choice. A model being able to suggest an action does not mean an application has implemented or authorized it.
Read model profiles · Browse applications and runtimes · Understand agent options
Where the software runs changes what you must manage.
A hosted service runs the model or agent on infrastructure operated by a provider. You usually need an eligible account and may encounter plan allowances or usage charges. Downloadable model files need a runtime: software that loads the model and performs inference, the process of generating an output. Ollama, llama.cpp and vLLM are examples of runtimes with different setup and serving goals.
Installing an application on your computer does not establish that all its processing stays there. A local interface can call a cloud model, an external search service or a remote embedding provider. To assess a private workflow, follow the whole path from the input through each connected service and into storage. Local processing also means taking responsibility for compatible hardware, downloads, updates and backups.
How Ollama runs models · Choosing a runtime for your workload · A real local document archive
Start with an input and an output you can check.
Choose a small task whose result you can recognize: turn a long task into a checklist, describe a photograph, or make a scan searchable. Read a real implementation to see how the builder connected the pieces. The case studies identify the actual stack and historical versions where disclosed; they do not silently replace those versions with the newest model.
For your own first trial, use material you are allowed to process, describe the desired output and decide how you will check it. Compare extracted text with the scan, a summary with the source, or a code change with a test and reviewed diff. This is a proposed evaluation method, not a claim that any particular product will pass. Our practice workshops supply separate fictional files so you can try a technique without starting with private data.
Task breakdown with Goblin Tools · Photo descriptions with Be My AI · Practice files and tutorials
Discovery lists show attention, not a verdict.
A repository is a project folder containing code, documentation and sometimes model files. GitHub hosts software repositories; Hugging Face also hosts model repositories and their model cards, which describe intended uses, access and requirements. Opening a repository is a research step, not an instruction to install it.
GitHub stars express interest in a project. They are not downloads, active users or evidence of successful deployments. The daily and weekly Trending snapshots preserve GitHub’s page order across all topics and languages, so position is not simply a sort of the visible period-star count. Our separate most-starred and recently-created lists use a defined sample of six topics.
Hugging Face Trending uses a provider-defined score, not a growth percentage or a quality test. Its most-downloaded view counts query-file requests over the trailing 30 days, not unique people. It includes retrieval and other model types as well as chat models. Every view retains its capture time; this edition does not refresh itself automatically.
Explore model snapshots · Explore repository snapshots · Read the exact populations and methods
Technical terms you will see in profiles
- Model weights
- The learned numerical parameters distributed as model files. Access to weights does not by itself settle commercial-use terms or provide a complete application.
- API and endpoint
- An API is an interface that software uses to request a service. An endpoint is the address for those requests. API usage can have separate billing from an assistant subscription.
- Context
- The material available to a model while producing a response, such as a prompt, supplied documents and conversation history. A large stated context limit does not guarantee perfect recall or reasoning.
- Embeddings and retrieval
- Embeddings represent inputs numerically so software can find related material. A document assistant can retrieve relevant passages before asking a model to answer; people still need to verify the cited passage supports the answer.
- Quantization
- A way to represent model numbers using fewer bits, often to reduce storage or memory needs. The exact representation and runtime affect what can run and how well it performs.
- MCP and agent frameworks
- Model Context Protocol is an interface for exposing tools to compatible clients. An agent framework helps developers connect models, state and tools. Neither supplies a finished, reliable workflow on its own.
This orientation explains concepts used in the linked, source-backed profiles. Individual product facts retain their own research dates. It is not a fresh product benchmark or a new metric capture.