Download the models in advance
Download your model before the workshop. First find your tier on Choose a model size for your hardware. The tier depends on your computer’s memory. With 16 GB or more you use the 8B model, and with 8 GB you use the 4B model. Then follow the tab for the tool you installed, Ollama or LM Studio.
16 GB or more (8B)
- In the Ollama app, open the model picker and type exactly
nishad/apertus:8b. - Select the entry with the download arrow and send any message. The download bar appears in the chat.
Typing apertus alone shows “No models found”. Always type the full tag, which is the model’s complete name.
Or run this in a terminal:
ollama pull nishad/apertus:8b8 GB (4B)
Do the same with nishad/apertus:4b (2.4 GB), or run ollama pull nishad/apertus:4b in a terminal.
Optional preview
If you want to compare, nishad/apertus:8b-v1.5-preview is the newer Apertus v1.5, converted for local use by a community member. Nothing in the workshop needs it.
Embeddings for Part 2
Part 2 also needs an embedding model. This second model does not write answers. It turns text into lists of numbers that represent the text’s meaning, and Part 2 uses those numbers for searching. This step works in the terminal only:
ollama pull qwen3-embedding:0.6bIt is 639 MB.
Sizes and budget
| Item | Size |
|---|---|
| Ollama install (Windows) | about 4 GB |
nishad/apertus:8b |
5.1 GB |
nishad/apertus:4b |
2.4 GB |
qwen3-embedding:0.6b |
639 MB |
| Python environment | about 0.5 GB |
The total stays well under the 20 GB of free disk space we ask for. On a 20 Mbit/s connection, the 8B model takes about 35 minutes to download.
Two index files for Part 2
Notebook 03 needs two vector index files, stw-de-labels.sqlite (85.7 MB) and unesco-en-labels.sqlite (34.4 MB). A vector index stores the lists of numbers that an embedding model makes from each text. It can then find texts with similar meanings quickly. Building these two yourself is slow on most laptops, so download the prebuilt files instead. The download link is in step 4 of Before you start. Put them in the runs/ folder of the notebooks repository. To confirm that the files arrived complete and unchanged, use the checksums and the command on the environment page.
Offline alternative
Show the offline options
Copy the manifests and blobs folders from ~/.ollama/models on a colleague’s machine into your own models folder. Or open Settings, then Model location, and select a USB stick that holds a copy of those folders.
Another option is to create the tag yourself from a GGUF file, the single-file format used for local models. Use the Modelfile from the models/ folder of the notebooks repository:
ollama create nishad/apertus:8b -f ModelfileVerify
Open a chat with the model you downloaded and send this message:
Reply with the word OK and the model name you think you are.
Read the answer. Then, if you use Ollama, run this in a terminal to list your models:
ollama listYour tag should be in the list. In LM Studio, the model appears under My Models instead.
Digests and sources
A digest is a code computed from the model’s contents; it identifies the exact file you downloaded, and ollama list shows its first 12 characters. Two tags with the same digest are the same model.
| Tag | Source GGUF | GGUF sha256 | Digest (ollama list) |
|---|---|---|---|
nishad/apertus:8b (also :latest) |
unsloth/Apertus-8B-Instruct-2509-GGUF, Q4_K_M |
87e7f0eb5a1d33e1a8ea4e7fdd2363764e0465061dd92962e0073aec31a7944d |
4a85a6fea0b8 |
nishad/apertus:4b |
jondale/Apertus-v1.1-4B-Instruct-GGUF, Q4_K_M |
not recorded | 87ca69a79aed |
nishad/apertus:8b-v1.5-preview |
Colby/apertus-v1.5-8b-text-Q4_K_M-GGUF, Q4_K_M |
a037df8d87ff6caacee794ee85f55342f2152e0d359b7389033300c3bee5f299 |
6f25f6ba0b62 |