Model Summary

Apertus 1.5 70B is Switzerland's latest fully open multilingual large language model developed by the Swiss AI Initiative. It is designed to advance transparent, compliant, and high-performance AI while remaining fully open throughout the entire training pipeline.

The model supports more than 1,000 languages, natively covers 1,811 languages, provides an extended 262,144-token context window, and is trained exclusively on compliant, openly available datasets. Apertus 1.5 delivers performance competitive with models trained behind closed doors while maintaining complete transparency of its weights, datasets, and training methodology.

The model is a decoder-only Transformer pretrained on 17 trillion tokens using a staged curriculum of web, code, mathematics, and domain-specific datasets. It introduces the xIELU activation function together with the AdEMAMix optimizer. Post-training combines supervised fine-tuning with QRPO, DPO, and RLVR alignment techniques to improve reasoning quality, instruction following, and conversational behaviour.

Key Features

• Swiss AI Initiative – Developed by the Swiss AI Initiative to provide a fully open, transparent, and compliant foundation model.
• Fully Open – Open model weights, open training data, complete training recipes, and publicly documented training methodology.
• Massively Multilingual – Native support for 1,811 languages, enabling broad multilingual understanding and generation.
• Compliance First – Trained using fully compliant datasets while respecting retrospective opt-out requests and reducing memorization through the Goldfish loss methodology.
• Long Context – Supports context windows of up to 262,144 tokens, enabling large document analysis and extended conversations.
• Tool Calling Support – Supports agentic workflows and tool use when deployed with compatible inference frameworks such as recent versions of vLLM.
• Open Deployment Ecosystem – Compatible with Transformers, vLLM, SGLang, MLX and GGUF deployment formats.

What's New in Apertus 1.5

Compared to the original Apertus release, version 1.5 introduces several major improvements:

• Continued Pretraining – The training corpus has been expanded with an additional 2 trillion high-quality tokens, with increased focus on public-interest domains.
• Advanced Reasoning – Improved logical reasoning enables better performance on complex, multi-step tasks.
• Improved Instruction Following – Better adherence to user instructions results in more accurate, predictable, and reliable responses.
• Longer Context Window – Increased from the original release to 262K tokens, enabling significantly larger inputs and documents.

Use Cases

Apertus 1.5 70B is particularly well suited for:

• Enterprise conversational AI
• Long-document summarization and analysis
• Multilingual assistants supporting more than 1,800 languages
• Government and public-sector AI workloads requiring transparency and compliance
• Retrieval-Augmented Generation (RAG)
• Agentic applications using tool calling
• Research and academic applications requiring fully open models
• Organizations preferring Swiss-developed, transparent AI technologies

How to use

Python

import os
from openaiimport OpenAI

client= openai.OpenAI(
    api_key=os.getenv("YOUR_API_KEY"),
    base_url="https://api.swisscom.com/products/swiss-ai-weeks/apertus-1.5-70b/v1",
)

stream= client.chat.completions.create(
    model="swiss-ai/Apertus-v1.5-70B",
    messages=[
        {"role":"user","content":"What is the most beautiful city in Switzerland?"}
    ],
    temperature=0.8,
    top_p=0.9,
    stream=True
)

for chunkin stream:
    print(chunk)
    print(chunk.choices[0].delta.content)
    print("****************")

curl

curl -X POST https://api.swisscom.com/products/swiss-ai-weeks/apertus-1.5-70b/v1/chat/completions \
  -H "Authorization: Bearer $SWISSCOM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "swiss-ai/Apertus-v1.5-70B",
    "messages": [
      {"role": "user", "content": "Wie geht es dir?"}
    ]
  }'

Context Window

Apertus 1.5 supports a context window of up to 262,144 tokens, representing a four-fold increase over the original Apertus release. This enables long-document processing, Retrieval-Augmented Generation (RAG), and complex multi-turn conversations.

Limits

• 5 Requests/s
• Input Tokens Budget 10000000
• Output Tokens Budget 2500000
• Authorization Bearer expires in 60 minutes

Hot to get your free API key

Technical Details

• Architecture: Transformer decoder
• Pretraining tokens: 15T
• Precision: bfloat16

Open resources

All elements used in the training process are made openly available

• Training data reconstruction scripts: github.com/swiss-ai/pretrain-data
• The training intermediate checkpoints are available on the different branches of this same repository