AI Tools

What Is the Apertus Open Source AI Model and Should You Use It?

Published 2026-06-23 · Source-based research, not a hands-on review

Apertus is a fully open-source large language model from Switzerland's top research institutions, released under Apache 2.0 with open training data, code, and weights. It comes in 8B and 70B parameter sizes, was trained on 15 trillion tokens across 1,500-plus languages, and is designed for EU AI Act compliance from day one.

Key takeaways

The short version

Apertus is the first major foundation model that is fully open — not just open weights, but open training data, code, and alignment methods, all under the permissive Apache 2.0 license.
Built by ETH Zurich, EPFL, and the Swiss National Supercomputing Centre as a public-interest project, not a commercial AI lab.
Trained on over 15 trillion tokens across 1,500+ languages (40% non-English), with built-in EU AI Act and GDPR compliance measures including PII removal and opt-out mechanisms.
Available in 8B and 70B parameter sizes, plus a Mini collection (0.5B–4B) for research; deployable via LM Studio, vLLM, SGlang, and community GGUF/MLX builds.

What Is the Apertus Open Source AI Model?

Apertus is a large language model developed by the Swiss AI Initiative, a collaboration between ETH Zurich, EPFL, and the Swiss National Supercomputing Centre (CSCS). The project began in late 2023 and was trained on the Alps supercomputer using over 10,000 NVIDIA GH200 Grace-Hopper chips, per the project's official documentation.

The name comes from the Latin word for "open," and that is the model's defining trait: Apertus publishes its training data, source code, model weights, and alignment methodology in full, not just the final weights. It is released under the Apache 2.0 license, which permits both research and commercial use without royalties.

The model was trained on more than 15 trillion tokens with a 40% non-English data mix spanning 1,500-plus languages. It ships in two main sizes — 8 billion and 70 billion parameters — with both base and instruction-tuned variants available on Hugging Face under the swiss-ai organization.

What Makes Apertus Different from Other Open Source AI Models?

Most "open source" AI models — including Meta's Llama and Mistral — release model weights but keep training data and code closed. Apertus takes a different approach. Per the Swiss AI Initiative, it is fully open: training data, code, weights, methods, and alignment principles are all documented and reproducible.

Three things set it apart from models like Llama 4 and Mistral 3:

Complete openness. You can inspect and reconstruct the training pipeline, not just download a weights file. This matters for researchers verifying results and for enterprises that need to audit model provenance.
EU AI Act compliance built in. Apertus was designed to meet EU AI Act requirements from day one. It includes opt-out mechanisms, PII removal from training data, and memorization prevention. The project publishes a public compliance summary and follows the EU Code of Practice for general-purpose AI.
Public-institution backing. Unlike models from commercial labs (OpenAI, Google, Anthropic) or US startups (Meta, Mistral), Apertus was built by Swiss public universities with no corporate owner. Swisscom is a strategic partner, but the project is governed by the Swiss AI Charter, an internal constitution prioritizing transparency and public-interest use.

Apertus Model Sizes and Performance

Apertus is available in several sizes, all under the same Apache 2.0 license. The official Hugging Face organization (swiss-ai) hosts base and instruct versions:

Apertus 8B — 8 billion parameters. Competitive with similarly sized open models. Suitable for consumer GPUs and local inference via LM Studio or GGUF.
Apertus 70B — 70 billion parameters. The flagship model, requiring more compute (recommended: vLLM or SGlang for server deployment).
Apertus Mini — A set of 16 small models (0.5B, 1.5B, 4B parameters) released June 15, 2026 to demonstrate distillation and quantization techniques.
Apertus 1.5 — An 8B + 70B combination model, listed as "coming soon" on the official site.

The project's technical report has been accepted at ACL 2026, the leading NLP conference, and includes detailed evaluation results. The model is described as "competitive with top open models at an equivalent scale" per the official website — but independent benchmarks comparing Apertus directly to Llama 4 or Mistral 3 on standard tasks like MMLU and HumanEval are not yet widely published.

Apertus vs Llama vs Mistral: How They Compare

If you are evaluating open-source models for a project, here is how Apertus stacks up against the two most popular open-weight alternatives. All three models permit commercial use, but their openness and compliance profiles differ significantly.

Who Should Use Apertus (and Who Shouldn't)

Apertus is not a ChatGPT-style consumer product. The project's FAQ states clearly: "Apertus is not a consumer product: this is foundational infrastructure." It is designed for organizations that need an open, auditable, EU-compliant model they can fine-tune and deploy on their own infrastructure.

Good fit if you:

Need an AI model you can inspect end-to-end — training data, code, and weights — for compliance or research.
Operate in the EU and need GDPR and AI Act alignment without building it yourself.
Work with multilingual content and need strong non-English performance out of the box (1,500+ languages in training).
Want a permissively licensed (Apache 2.0) foundation model for commercial fine-tuning or product development.
Prefer a model from public, non-commercial institutions with transparent governance.

Look elsewhere if you:

Need a ready-to-use chat application — Apertus is model infrastructure, not a product. You can try it via the community chat demo at chat.publicai.co, but official hosted access is not yet available.
Need the highest benchmark scores today — Apertus is still early in its lifecycle and independent benchmarks are limited. Llama 4 and Mistral 3 have more published evaluation data at scale.
Need an ecosystem of plug-and-play tools, APIs, and third-party integrations — the Apertus ecosystem is new compared to the mature tooling around Llama and Mistral.

How to Get Started with the Apertus Open Source AI Model

You can download and run Apertus today through several channels, all documented on the official site:

Hugging Face — The swiss-ai organization hosts all model weights. Download base or instruct versions directly.
LM Studio — For quick local inference on Mac, Windows, or Linux. Apertus is listed in the LM Studio model catalog.
vLLM and SGlang — Recommended for self-hosted inference servers. SGlang is suggested for advanced scaling needs.
Community builds — GGUF and MLX format conversions are available from community publishers. The Apertus team notes these are unofficial and recommends verifying provenance.
Fine-tuning — The project provides a fine-tuning guide in its documentation. Since the training data and code are open, you can also reproduce or extend the training pipeline yourself.

For Python developers, the project publishes the Apertus Format library, a Python package for the model's custom chat format. Training code, evaluation scripts, and model artifacts are available on the Swiss AI GitHub organization.

At a glance

Feature	Apertus	Llama 4 (Meta)	Mistral 3
License	Apache 2.0	Llama 4 Community License	Mistral Research License / Commercial
Open training data	Yes — fully published	No — not released	No — not released
Open source code	Yes — GitHub (swiss-ai)	Partial — inference code only	Partial — inference code only
EU AI Act ready	Yes — built-in compliance	Not explicitly designed for	Not explicitly designed for
Developer	Swiss AI Initiative (ETH/EPFL/CSCS)	Meta AI	Mistral AI (France)
Training languages	1,500+	~30 (primarily English)	~12 (primarily English/French)
Model sizes	0.5B, 1.5B, 4B, 8B, 70B	8B, 70B, 400B	7B, small, medium, large
Commercial use	Yes (Apache 2.0)	Yes (with restrictions)	Yes (paid tier for large models)

Apertus vs Llama vs Mistral at a glance

FAQ

Is the Apertus AI model free for commercial use?

Yes. Apertus is released under the Apache 2.0 license, which permits both research and commercial use without royalties or restrictions. You can fine-tune it on proprietary data and deploy it in commercial products, provided you comply with the license terms.

Can I run the Apertus 70B model on my own computer?

The 70B model requires significant GPU memory — typically multiple high-end GPUs or a server with at least 140 GB of VRAM at full precision. The 8B model is much more practical for local use: it runs on consumer GPUs with 16-24 GB VRAM via LM Studio or GGUF quantized formats. The Mini models (0.5B-4B) run comfortably on most modern laptops.

How does Apertus compare to Claude or ChatGPT?

Apertus is a foundation model you download and run yourself, not a hosted chat service like Claude (Anthropic) or ChatGPT (OpenAI). It competes more directly with open-weight models like Llama and Mistral. As of June 2026, independent benchmarks comparing Apertus to commercial chat models are limited, and Apertus is explicitly positioned as research infrastructure rather than a consumer AI assistant.