Instructions to use swiss-ai/Apertus-70B-Instruct-2509 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use swiss-ai/Apertus-70B-Instruct-2509 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="swiss-ai/Apertus-70B-Instruct-2509")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("swiss-ai/Apertus-70B-Instruct-2509")
model = AutoModelForCausalLM.from_pretrained("swiss-ai/Apertus-70B-Instruct-2509", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Inference
HuggingChat
Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use swiss-ai/Apertus-70B-Instruct-2509 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "swiss-ai/Apertus-70B-Instruct-2509"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "swiss-ai/Apertus-70B-Instruct-2509",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/swiss-ai/Apertus-70B-Instruct-2509

SGLang

How to use swiss-ai/Apertus-70B-Instruct-2509 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "swiss-ai/Apertus-70B-Instruct-2509" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "swiss-ai/Apertus-70B-Instruct-2509",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "swiss-ai/Apertus-70B-Instruct-2509" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "swiss-ai/Apertus-70B-Instruct-2509",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use swiss-ai/Apertus-70B-Instruct-2509 with Docker Model Runner:
```
docker model run hf.co/swiss-ai/Apertus-70B-Instruct-2509
```

Are <|inner_prefix|>...<|inner_suffix|> thoughts generated by released Instruct checkpoints?

#18

by pd95 - opened May 17

Discussion

pd95

May 17

Hi, I am experimenting with running Apertus Instruct through Ollama's MLX runner and have a question about thinking/deliberation output.

The official chat template supports:

enable_thinking=True
Deliberation: enabled
assistant thoughts blocks rendered as <|inner_prefix|>...<|inner_suffix|>

I wired this locally so that think: true enables Deliberation: enabled, and the parser extracts <|inner_prefix|>...<|inner_suffix|> into a thinking field.

However, in smoke tests with Apertus-8B-Instruct-2509 and Apertus-70B-Instruct-2509, the models answer normally but do not emit any <|inner_prefix|>...<|inner_suffix|> spans. The API path appears to work, but
thinking remains empty.

Is this expected behavior for the released Instruct checkpoints?

Concretely:

Are Apertus Instruct models expected to generate visible/parseable thoughts blocks when enable_thinking=True?
Are there recommended prompts or decoding settings that induce this behavior?
Is Deliberation: enabled intended only as an internal instruction, with the model usually returning only the final response?
Should downstream integrations advertise this as a thinking-capable model, or only as supporting the Apertus thinking format for replay/history?

Thanks!

loleg

Swiss AI Initiative org May 25

This is a known issue of our limited support for instruction following in the first release. For now, I would not label the current version of Apertus as a thinking model. Our team is working on making sure the next version of Apertus improves support here, and uses thinking special tokens more frequently. Thanks for your interest, and please stay tuned!

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment