Model literacy

Where a model runs. What you can change.

A model, the app around it, and the service hosting it are different parts of a system. Check the exact version and configuration before making a choice.

Hosted, local, and mixed setups

Hosted

The provider runs the model on remote computers. Your request reaches that service. Check account terms, retention, sharing, access controls, and pricing. The app may also connect to other tools.

Local

Inference runs on your device or a server you manage. You take responsibility for installation, memory, updates, and access. Check whether the surrounding app still uses remote services or stores sensitive logs.

Mixed

A local interface can call a cloud model, retrieve web content, or use remote tools. The location of the interface does not establish the location of processing. Trace the full data path.

Concrete example: Ollama documents both local use and cloud features, with a local-only setting that disables its cloud models and web search. This does not establish the behavior of every app connected to Ollama. Read Ollama’s FAQ ↗

Open weights, open source, and licenses

What does “open weights” give you?

Weights are the learned numerical parameters. Downloadable weights can let you run a model with compatible software. Check the license for permission to use, modify, or share them. Availability alone does not supply training code or data information.

What does “open source AI” mean?

Under the Open Source Initiative’s definition 1.0, the system must support freedoms to use, study, modify, and share. It also needs data information, code, and parameters in forms suitable for modification. “Open” is used differently by different publishers; identify the definition and terms being claimed. OSI definition ↗

What should I read before using a model?

Read the exact model card and license, including version, intended uses, limitations, evaluation, and restrictions. Check the software license separately. A model family name does not establish every release’s license or suitability. Hugging Face model-card guide ↗

Read an Arena ranking carefully

Standard Arena rankings aggregate preferences from head-to-head comparisons. Inspect the task category, model version, vote count, confidence interval, and rank spread. A relative score is not a percentage of correct answers. Arena’s leaderboard guide ↗

Arena’s factuality view combines preferences with checks of web-verifiable claims. Read which view is selected and how it is scored. This still does not certify an answer, prove privacy, or validate use in your classroom. Factuality methodology ↗

Why hardware requirements vary

Model weights need storage and working memory. Runtime overhead, context, and concurrent requests need additional memory. Ollama supports CPU, GPU, and mixed loading; compatibility depends on hardware and software. Hardware support ↗ · Memory and concurrency FAQ ↗

Quantization stores weights at lower precision to reduce memory. Quality and speed depend on the method, model, and hardware. Test the resulting model on your task. Hugging Face quantization guide ↗

Explore weight storage

A simplified calculation: billions of stored parameters × bits per weight ÷ 8 = decimal GB of raw weights.

Stored parameters

Illustrative precision

Weight-only arithmetic, not a device recommendation or actual download size. Additional memory is required for runtime, context, and quantization metadata. Use total stored parameters rather than an advertised count of active parameters. This illustration covers inference; training needs a different memory estimate.

A small comparison before committing

Try the same representative tasks and check accuracy, corrections needed, latency, accessibility, and cost. Include privacy and license requirements in the comparison. Use public or fictional test material. A larger model or higher rank does not settle suitability.

Explore tools and model providers in the library → · Check an AI claim →

Sources checked October 11, 2026. Definitions and the comparison prompts are this hub’s synthesis. OSI supplies a definition; Ollama and Hugging Face document particular software; Arena describes its evaluations. No model is installed or called by this guide.