How to choose a model

Choose for a task

Write the job in one sentence. “Extract dates and owners from short English meeting notes” is a job you can test. Decide which languages, file types, and response formats are required. For ordinary chat, start with an instruction-tuned model. Use a vision-capable model when the input includes images, and an embedding model when you need semantic search.

Read a model card

• Training and purpose. Check whether the model is a base or instruction-tuned variant and what use cases it was evaluated on.
• Compatibility. Check architecture, file format, quantization, runtime support, and available memory. The same model family can have different sizes, features, and context limits.
• Permission. Read the license for the exact model and weights you plan to use. Open weights make a download possible. The license describes conditions for use, redistribution, or derivatives.
• Evidence. Look for evaluations that resemble your task and sample your own inputs before relying on it.

Try it

Open the linked Qwen model card. Identify its instruction-tuned variant, a GGUF quantization, and the license. Compare the upstream GGUF file listing with the Ollama package for the same size class. File sizes and packaging can differ across distributions.

Then create three short test inputs. For meeting-note extraction, try a note with an explicit due date, a note with no date, and a note with two dates. Pick the smallest model that passes the examples at a speed you can use, then build a larger test set.

Check your understanding

Two downloads share a family name, but one says “Base” and one says “Instruct.” Which is the starting point for a helpdesk chat?

Start with the instruction-tuned variant and test it against examples of your actual helpdesk requests.

Sources

• Hugging Face Hub · Model cards
• Qwen · Qwen2.5-1.5B-Instruct GGUF
• Ollama Library · qwen2.5:1.5b