Start here · Step 6 of 7
Step 6 of 7 · about 8 minutes
First chat
Use a tag that exists on the official library page for the model you picked. For Ollama, after you have chosen Qwen3-8B, the shape is ollama run qwen3:8b. Check the model page here before you treat any other tag as official.
In LM Studio, open the official app, download the same model from its catalog, and start a chat. I am not inventing catalog IDs on this lesson. If a vendor card publishes an lms get command, the matching model page may repeat it.
For llama.cpp, use the official GGUF repo when the publisher has one, and follow that card’s own command.
What to try: a short question, then a follow-up that needs the first answer, so you can see that context is real. Qwen3 has a thinking mode. Turn it off for a quicker first reply. The Qwen3 rows I measured used think off.
If it is slow, that may be normal on a CPU. I am not going to tell you it “should be” some number of tokens per second. The hardware lesson explains why. The benchmark table is only the machines I have actually timed.
If it fails
Go to It will not run. The short list, which does not cover every machine:
- The runner is not installed, or the tag or filename is wrong.
- The file does not fit in memory. Compare the error with the computed total on the model page, and pick a smaller model or a shorter context.
- The GPU is not seen, so the run fell back to the CPU or stopped.