---

license: apache-2.0

base_model: swiss-ai/Apertus-8B-Instruct-2509

base_model_relation: quantized

pipeline_tag: text-generation

library_name: gguf

tags:

- gguf

- llama.cpp

- apertus

- swiss-ai

- multilingual

language:

- multilingual

---

Apertus 8B Instruct 2509 - GGUF

Quantized from swiss-ai/Apertus-8B-Instruct-2509.

All credit to Swiss AI - this is a format

conversion, with the chat template tweaked to suit llama.cpp.

Chat template

llama.cpp hands the template tool arguments as a mapping rather than a string,

so we wrapped them before they go into the JSON:

-{{ '{"' + tool_call.name + '": ' + tool_call.arguments + '}' }}
+{{ '{"' + tool_call.name + '": ' + (tool_call.arguments if tool_call.arguments is string else tool_call.arguments|tojson) + '}' }}

Same again on the function.arguments line below it. That is the whole change.

Files

| | Size | |

|---|---|---|

| apertus-8b-instruct-2509-q4_k_m.gguf | 5.1 GB | the only one so far, ask if you want Q8_0 |

Running it

llama-server -m apertus-8b-instruct-2509-q4_k_m.gguf --ctx-size 8192 --reasoning off

It is a reasoning model. Left on, a small token budget goes entirely on

thinking and you get an empty content field back.

Made with

llama.cpp b10223, --convert then --quantize out of the :full image,

template tweaked before converting.