Each language model on this site carries task tags: chat, coding, reasoning, agentic and RAG. They say what the model is built and evaluated for, not what it is limited to.
| Tag | Means | Look for |
|---|---|---|
| Chat | Follows instructions and answers in a conversation. | An instruct model with a good grade that Runs well. |
| Coding | Writes and edits code. | The largest that runs well, at Q4_K_M or better (guide). |
| Reasoning | Thinks step by step before answering. | Speed and context to spare (guide). |
| Agentic | Calls tools and carries out multi-step tasks. | Reliable tool use and a long context. |
| RAG | Answers from supplied documents. | Good at sticking to the text it is given (guide). |
A simple way to shortlist
- Start from your hardware so the list is already filtered to what fits: the Model Finder on the home page, or press Detect on My hardware.
- Keep only “Runs great” or “Runs well”. A slightly stronger model that runs slowly is usually worse in practice (ratings).
- Sort by grade and filter by the task you care about.
- Take three from different families and try them on your own prompts. A model’s style matters as much as its score, and families differ in tone, strictness and language coverage.
Size is not everything
Newer models are often better than older ones twice their size, and mixture-of-experts models can be quick for their quality (MoE). Compare the grade, not the parameter count. The tier list ranks models across the catalog. Check the license too if you plan to build a product on one (licenses).