Model guide
Every model the workshop uses or mentions, with the facts that decide whether it suits your machine. We made the measurements in October 2026 on Ollama 0.35.1 and repeated them on 0.40.4 with the same counts. To choose a model for your memory size, start with Choose a model size for your hardware.
Apertus models
Apertus is the fully open model family from the Swiss AI Initiative (EPFL, ETH Zurich, CSCS). The workshop tags below are the names you give Ollama to download each model. The facilitator built them from GGUF files (a single-file model format) shared by the community. The Modelfiles that define them are in the notebooks repository under models/. All three tags set temperature 0.8 and top_p 0.9.
nishad/apertus:8b
The reference model for both parts. Also available as nishad/apertus:latest.
| Field | Value |
|---|---|
| Model | Apertus-8B-Instruct-2509, Unsloth Q4_K_M |
| Parameters | 8 billion |
| Size | 5.06 GB |
| Memory in use | about 5.3 GB at 4,096 and 5.8 GB at 8,192 context |
| Context | 65,536 tokens supported; 8,192 set in the tag |
| License | Apache-2.0 plus the Apertus Acceptable Use Policy |
| Openness tier | Fully open (tier 3) |
| Digest | 4a85a6fea0b8 |
| Source | hf.co/unsloth/Apertus-8B-Instruct-2509-GGUF:Q4_K_M |
| GGUF sha256 | 87e7f0eb5a1d33e1a8ea4e7fdd2363764e0465061dd92962e0073aec31a7944d |
It is good at extracting fields from catalog-style text, at German, French and Italian, at structured output, and at calling tools when the prompt is in English. It is weak at code, math and following instructions (a score of 71.7 on IFEval, a benchmark for instruction following). When a field is missing, it sometimes fills it from a neighboring field, for example by putting the publisher in the place field (see notebook 02). A chat template is the text pattern that wraps each message before it reaches the model. The tag uses the chat template stored in the GGUF file itself, with Unsloth’s fixes for tool arguments.
nishad/apertus:4b
For machines with 8 GB of memory.
| Field | Value |
|---|---|
| Model | Apertus v1.1 4B Instruct, Q4_K_M |
| Parameters | 4 billion |
| Size | 2.43 GB |
| Memory in use | about 2.8 GB |
| Context | 4,096 tokens (hard limit) |
| License | Apache-2.0 weights, with the Apertus LLM Acceptable Use Policy |
| Openness tier | Fully open (tier 3) |
| Digest | 87ca69a79aed |
| Source | hf.co/jondale/Apertus-v1.1-4B-Instruct-GGUF:Q4_K_M |
| GGUF sha256 | checksum not recorded |
The tag uses a wrapper template written for the workshop. It presents tools to the model in Apertus’s TypeScript-style format, so Ollama can read the model’s tool calls. Without it, tool requests never finish. Quality is clearly below the 8B. In our tests it could not recall a fact from the previous turn, and it could not find one fact hidden in a 2,300-token text.
nishad/apertus:8b-v1.5-preview
An optional preview for people with 16 GB or more who want to compare.
| Field | Value |
|---|---|
| Model | Apertus v1.5 8B, text only, Q4_K_M |
| Parameters | 8 billion |
| Size | 5.06 GB |
| Memory in use | as the 8B |
| Context | 8,192 set in the tag; the file’s metadata says 262k |
| License | Apache-2.0, with the Apertus Acceptable Use Policy v1.5 (14 July 2026) |
| Openness tier | Fully open (tier 3) for Apertus 1.5 itself |
| Digest | 6f25f6ba0b62 |
| Source | hf.co/Colby/apertus-v1.5-8b-text-Q4_K_M-GGUF:Q4_K_M |
| GGUF sha256 | a037df8d87ff6caacee794ee85f55342f2152e0d359b7389033300c3bee5f299 |
This is an unofficial community conversion. The original v1.5 model handles images and audio as well as text, and its download is gated (you must request access). Someone extracted only the text part; the vision and audio parts are removed, and nobody has checked that the result matches the original. Thinking is not available in Ollama. Because the file’s metadata says the context is 262k, a Mac with a lot of memory would reserve far more memory than you expect. Set the context size yourself; the workshop tag already does. In testing it scored slightly better on tool calls than the 8B.
Verification of the tags
We ran these checks on Ollama 0.35.1 and repeated them on 0.40.4 on 15 October 2026, with identical counts both times. We ran each model once on each Ollama version, through the OpenAI-compatible endpoint (Ollama’s local server, which accepts requests in the OpenAI API format).
| Tag | Tool prompts with a structured call | Controls clean | Multi-turn loops | Structured output |
|---|---|---|---|---|
nishad/apertus:8b |
16 of 20 | 10 of 10 | 5 of 5 | 5 of 6 |
nishad/apertus:4b |
11 of 20 | 10 of 10 | 3 of 5 | 5 of 6 |
The one structured-output request that fails uses a schema with an empty enum (a list of allowed values with nothing in it), and the server rejects it with HTTP 400 as expected. The 4B fills a missing ISBN with the text "2 Bände"; the 8B correctly returns null.
Embedding models
Embedding models turn text into vectors (lists of numbers that represent meaning). Semantic search uses those vectors to match by meaning rather than by exact words (see Glossary). The workshop default is qwen3-embedding:0.6b. It had the best result among the openly licensed models on a gold set (a test set with known correct answers) built for this workshop. The set has 50 queries over the STW Thesaurus for Economics and the UNESCO Thesaurus. We ran the evaluation in October 2026.
| Model | License | Size | Dimensions | Context | Prefixes | recall@1 | recall@5 |
|---|---|---|---|---|---|---|---|
qwen3-embedding:0.6b (default) |
Apache-2.0 | 639 MB | 1,024 | 32k tokens | instruction prefix on queries, added by the notebooks’ helper | 0.62 | 0.82 |
embeddinggemma |
Google’s Gemma Terms of Use, not an open-source license | 621 MB | 768 | not recorded here | task-string prefixes | 0.68 | 0.92 |
bge-m3 |
MIT | 1.2 GB | 1,024 | 8k tokens | none | 0.34 | 0.48 |
embeddinggemma scored higher, but its license is not an open-source one, so we list it as an alternative rather than the default. bge-m3 has a permissive license and its fine-tuning data is released, but it scored lowest on this gold set. How you build the index matters as much as which model you choose; notebook 03 shows the difference.
Apertus has no embedding model; the workshop uses qwen3-embedding:0.6b beside it.
Small open-weight fallbacks
These are open weights only: the weights are downloadable and the license is Apache-2.0, but training data and code are not fully open. Use them only if no Apertus tag runs on your machine.
| Tag | License | Size | Note |
|---|---|---|---|
qwen3.5:4b |
Apache-2.0 | 3.3 GB | Thinking must be switched off |
granite4.1:3b |
Apache-2.0 | 2.1 GB | Smallest option |
Ollama commands
| Command | What it does |
|---|---|
ollama list |
Show downloaded models with their digests |
ollama pull NAME |
Download a model |
ollama rm NAME |
Delete a model |
ollama show NAME |
Show details, template and capabilities |
ollama ps |
Show models loaded in memory |
ollama stop NAME |
Unload a model from memory |
What may change after October 2026
We checked these facts in October 2026 and will check them again the week before the workshop. Each of the changes below would affect you:
- An official Apertus entry in the Ollama library, or an official GGUF from the Apertus team, would replace the community builds behind the workshop tags.
- Support for the v1.5 architecture in llama.cpp and Ollama would make the preview tag unnecessary.
- New releases in the swiss-ai collection on Hugging Face may add sizes or versions.