Switch to another model

A Tesla P100 has 16 GB, and one launch target with its 65,536-token context takes nearly all of it. To use another model, stop the one that is running, then run the other. This page switches from Qwen3.8-27B-PN3 to Qwen3.8-27B-UD-IQ3_XXS-Q3Floor (Q3Floor); the steps are the same in the other direction.

You need the launcher installed and a host that passes tensorless doctor. See Prepare a Linux host for a P100.

1. See which models your launcher knows

tensorless models

MODEL                                            CARD  STATUS    ENGINE         FITS  CACHED
iamMrDuncan/Qwen3.8-27B-PN3-GGUF                 p100  released  engine-v0.1.0  yes   yes
iamMrDuncan/Qwen3.8-27B-UD-IQ3_XXS-Q3Floor-GGUF  p100  released  engine-v0.1.0  yes   no

• FITS says whether a GPU on this host matches the recipe (yes, no, or unknown without a GPU).
• CACHED says whether the model file is already downloaded: yes, partial (an interrupted download) or no.

The launch targets, with their sizes and measured speeds, are in Launch targets.

2. Stop the running model

tensorless stop iamMrDuncan/Qwen3.8-27B-PN3-GGUF

If you started it in the foreground, Ctrl-C in its terminal does the same.

Caution

Don’t start a second model while the first holds the card. On the same port, the launcher refuses with cannot use port 8080. On another port, the second server runs out of GPU memory while it loads. On a host with several GPUs, you can run one model per card instead: see Pick a GPU on a host with several.

3. Run the other model

tensorless run iamMrDuncan/Qwen3.8-27B-UD-IQ3_XXS-Q3Floor-GGUF

The first time, this downloads the 11.76 GB model file and checks its sha256. The engine image is the same one both recipes use, so it is not pulled again.

Each recipe also defines a short name, which works anywhere a repo id does:

Short name	Repo id
pn3	iamMrDuncan/Qwen3.8-27B-PN3-GGUF
q3floor	iamMrDuncan/Qwen3.8-27B-UD-IQ3_XXS-Q3Floor-GGUF

tensorless stop pn3
tensorless run q3floor --detach

The repo id is not case-sensitive, as on the Hugging Face Hub.

Check that it worked

Ask the server which model it serves:

curl -s http://127.0.0.1:8080/v1/models

The answer lists iamMrDuncan/Qwen3.8-27B-UD-IQ3_XXS-Q3Floor-GGUF. Clients that name the model in their requests need the new repo id: see Connect an OpenAI-compatible client.