Swiss AI Initiative · 2026-07-24 · major
Apertus 1.5 — Switzerland's fully open 8B/70B goes multimodal
Swiss AI Initiative (ETH Zurich, EPFL, CSCS) released Apertus 1.5 in 8B and 70B sizes. Continued pretraining adds native image and audio input, an optional thinking mode, and a 262K-token context. Apache-2.0, weights and training data open.
Switzerland's fully open Apertus grows eyes and ears, keeps its Apache-2.0 promise.
Key specs
| Context window | 262K tokens |
|---|---|
| 8 b pretrain tokens | +4T |
| 70 b pretrain tokens | +2T |
Quick facts
| Makers | ETH Zurich, EPFL, CSCS (Swiss AI Initiative) |
|---|---|
| Model sizes | Apertus-v1.5-8B, Apertus-v1.5-70B |
| License | Apache 2.0 (research and commercial) |
| Modalities in | Image, audio, text |
| Context window | 262,144 tokens (4× Apertus 1.0) |
| Training compute | Alps supercomputer, CSCS Lugano |
| What's new | Multimodal input, thinking mode, better tool use |
What is it?
Apertus 1.5 is the new release of Switzerland's fully open large language model, built by ETH Zurich, EPFL, and the Swiss National Supercomputing Centre (CSCS). It ships in 8B and 70B sizes and, for the first time in the family, takes images and audio as input alongside text. Weights, training data, and training details are all public under Apache 2.0.
How does it work?
The team took the Apertus 1.0 checkpoints and kept training them on the Alps supercomputer in Lugano. The 8B model saw 4 trillion additional text and multimodal tokens; the 70B saw 2 trillion. Post-training added an optional thinking mode for harder reasoning, better instruction following, better tool use, and a 262,144-token context window, four times the size of Apertus 1.0.
Why does it matter?
Apertus 1.5 gives Europe a sovereign, fully open multimodal option at a size teams can actually self-host. Because everything is public — data, weights, training code — regulators, hospitals, and public agencies can audit the whole stack. EPFL is already using it to build MeditronFO, an open medical LLM framework, and the Ticino cantonal government uses Apertus for in-house document translation.
Who is it for?
European labs, public-sector teams, and researchers who need an auditable open-weight multimodal model.
Frequently asked questions
- Where can I download Apertus 1.5?
- Apertus 1.5 is on Hugging Face under the swiss-ai organization: swiss-ai/Apertus-v1.5-8B (9B parameters) and swiss-ai/Apertus-v1.5-70B (72B parameters). Both sit inside the swiss-ai/apertus-v15 collection with weights, tokenizer, and configs.
- Is Apertus 1.5 really fully open?
- Yes. Apertus 1.5 ships under Apache 2.0 with open weights, open training data references, and public training details, so researchers and companies can inspect, fine-tune, or deploy it commercially without a separate license request from ETH Zurich or EPFL.
- How is Apertus 1.5 different from Apertus 1.0?
- Apertus 1.5 is a continued pretraining of the 1.0 checkpoints. The 8B saw 4 trillion more tokens and the 70B saw 2 trillion more. It adds native image and audio input, an optional thinking mode for harder reasoning, and quadruples the context window to 262K tokens.
- Can Swiss researchers run Apertus 1.5 without their own GPUs?
- Yes. CSCS launched an inference endpoint for the Swiss academic community, hosted on the same Alps supercomputer in Lugano that trained the model. Everyone else can pull the weights from Hugging Face or wait for the third-party inference providers the team is onboarding.
- What is Apertus Mini?
- Apertus Mini is a companion suite of 16 compact models released alongside Apertus 1.5. The Swiss AI Initiative built them through distillation and quantization so smaller devices and edge deployments can run Apertus-family models without a data-centre GPU.
Try it
huggingface.co/swiss-ai/Apertus-v1.5-8B