SumizAI vs Apertus

What Apertus is

What Apertus is

Apertus is a fully open, multilingual large language model released on September 2, 2025 by EPFL, ETH Zurich and the Swiss National Supercomputing Centre (CSCS) — Switzerland's first large-scale open language model. The name, Latin for "open," reflects the project's philosophy: the entire development process — architecture, model weights, training data and training methods — is publicly available and fully documented, setting it apart from most commercial models that treat training data as a trade secret.

Key features

The model was trained on 15 trillion tokens spanning more than a thousand languages, with 40% of the training data being non-English — including languages so far underrepresented in large models, such as Swiss German and Romansh. Apertus ships in two sizes, 8 billion and 70 billion parameters, letting teams match the model to available compute. A follow-up release, Apertus 1.5, brought further infrastructure improvements, and the accompanying Apertus Mini package offers sixteen compressed variants showing how distillation and quantization can deliver reasonable quality on more modest hardware.

Who it's for

Apertus appeals mainly to researchers, academic institutions and companies that need full transparency into the training process — for auditing systemic bias, for instance, or verifying which data influenced the model's outputs. It also suits anyone working with languages poorly served by closed models, and public-sector teams in Europe for whom technological sovereignty and data-locality compliance are a priority.

Use cases

The model is available through strategic partner Swisscom, the Hugging Face platform, and the Public AI network, which in practice allows both local deployment and use of a hosted API. Research institutions use Apertus to analyze text in smaller European languages, while engineering teams build their own fine-tuned variants on top of it for specific tasks, from document classification to multilingual customer support.

Pricing and business model

Apertus is free and open-source under the Apache 2.0 license, permitting commercial use, modification and redistribution without licensing fees. The only cost is the infrastructure needed to run the model locally, or fees for hosted services from partners such as Swisscom.

Limitations and what to watch for

As a fully open model, Apertus doesn't come with a ready-made chat interface comparable to commercial assistants — running it requires either your own infrastructure or a partner hosting platform. The larger 70-billion-parameter variant needs significant GPU resources, which can be a barrier for smaller teams without access to a supercomputer.

AI notes from a model trained on over a thousand languages

For teams building their own ai notes tooling, it's worth knowing that Apertus handles many languages at once better than models trained predominantly on English — which matters when generating ai notes from conversations held in less common European languages, where other models often lose grammatical nuance.

The supercomputing infrastructure behind the model

Training ran on CSCS infrastructure, one of Europe's leading supercomputing centers, making Apertus one of the few fully open models at this scale trained entirely in Europe, without dependence on American or Chinese cloud infrastructure.

Multilingual by design, not as an afterthought

Unlike models that add support for minority languages as an extension of an already-existing, mostly-English corpus, Apertus's creators planned the training-data proportions from the start so that niche languages received real representation rather than merely a token presence in the dataset.

Open weights as a foundation for your own ai notes tooling

Teams building their own tools to turn AI conversations into ai notes increasingly reach for fully open-weight models instead of a closed provider's API, and Apertus gives them that option without the licensing restrictions typical of partially open models. Being able to run the whole ai notes generation process locally, without sending conversation content to outside servers, is often a non-negotiable requirement in regulated sectors like public administration or healthcare. For such teams, these notes produced with Apertus never leave infrastructure they control themselves, meaning every these notes entry falls under the same security regime as the rest of the organization's data instead of landing in an external model provider's cloud.

What's worth checking before choosing

Before deploying, check the current licensing documentation on apertus.ai, which model size (8B or 70B) matches your available compute, and whether a partner's hosted API meets your data-locality requirements.

Bottom line

Apertus is a rare example of a large language model where transparency matters as much as raw answer quality — a research project first, not a commercial product. Anyone looking for a ready-made tool that turns AI conversations into these notes without running their own infrastructure will find a more direct path elsewhere.

Comparison with SumizAI

Apertus and SumizAI operate on entirely different levels: Apertus is a raw language model requiring your own infrastructure or a hosting partner, while SumizAI is a finished note-taking app with a chat interface. SumizAI doesn't ship its own model — the user brings an API key for one of seven providers (Anthropic, OpenAI, Gemini, Groq, OpenRouter, a local Ollama, or their own server), so a self-hosted, Apertus-compatible endpoint could in principle power conversations inside SumizAI. SumizAI costs $1 a month after a seven-day trial with no card required, and answers worth keeping land automatically as Markdown files in a vault the user owns on their own disk, with no manual copying.

That is our reading of the record. For AlternativeTo's own wording, see its page for Apertus — the facts above were read on 2026-08-16.

Apertus in depth

A public-institution language model where openness is the point

Apertus is a large language model released by the Swiss AI Initiative, a collaboration between EPFL, ETH Zurich and the Swiss National Supercomputing Centre. It went public on 2 September 2025 under the Apache 2.0 licence. The name is Latin for open, and the choice is not decorative: what separates Apertus from most models described as open is that the openness extends past the weights to the training data, the training methods, the alignment principles and the code, all documented and intended to be reproducible. Plenty of models publish weights and call themselves open. Very few publish enough for an outside team to retrace how the thing was made, and that reproducibility is the specific contribution here.

Scale, languages and architecture

Two sizes were released initially, at 8 billion and 70 billion parameters — the model card for the larger instruct variant records 71 billion — with smaller variants at 0.5, 1.5 and 4 billion also published. Training consumed 15 trillion tokens. The multilingual claim is the headline number: over a thousand languages according to the launch materials, with the model card putting natively supported languages at 1,811, and roughly forty per cent of the training data non-English. That proportion is the reason the model performs on languages that usually get token coverage as an afterthought, including Swiss German and Romansh alongside German, French, Italian and English. Architecturally it is a decoder-only transformer using the xIELU activation function and trained with the AdEMAMix optimiser, with a default maximum context of 65,536 tokens. Training ran on 4,096 GH200 GPUs on the Alps supercomputer at the national supercomputing centre. The technical report was posted in September 2025 and presented at ACL 2026.

Data handling as a design constraint

The most consequential engineering decisions in Apertus are about what was allowed into the corpus. The project states that it uses only fully compliant and open training data, honours the opt-out signals of data owners including requests made retroactively after crawling, strips personally identifiable information, and applies techniques intended to prevent the model memorising and regurgitating its training material. There is a stated route for data protection and copyright removal requests. This is an unusual position to take, because respecting retroactive opt-outs costs data and therefore costs benchmark points. For an organisation in a European regulatory environment, and for anyone deploying a model where provenance of the training data is a compliance question rather than a preference, it is the difference between a model you can use and one you cannot.

How to actually use it

There are three routes. You can download the weights from the Hugging Face collection published by the swiss-ai organisation and run them yourself, which is the point of the Apache 2.0 licence — commercial use, modification and redistribution are all permitted without a separate agreement. You can reach it through Swisscom, the initiative's strategic partner, as a hosted service. Or you can use it free of charge through the Public AI Inference Utility, a non-profit that runs a chat interface and became an official inference provider on Hugging Face in September 2025, with deployment work carried out alongside AWS in Zurich and Intel. The chat front end is what the catalogue lists as the product's website, and it is where most people who are not running their own inference will meet the model.

What version 1.5 changed

The project did not stop at launch. Apertus 1.5 was announced on 24 July 2026 and is the release that changes what the model can be used for: it adds image understanding, so the model accepts visual input rather than text alone; it extends the context window by a factor of four; and it improves reasoning, tool use and instruction following. Tool use in particular matters for anyone considering the model as an agent backend, since that capability is usually where open models fall furthest behind commercial ones. The 1.5 release is available through Public AI, and the initiative continues to publish news, run an SME outreach programme in Switzerland and maintain the model as ongoing public infrastructure rather than a one-off publication.

What early users say

The catalogue holds a small body of feedback — an average of 3.8 out of 5 from five ratings, with twenty-nine likes and three written comments — and it is mixed in an informative way. On the positive side, one user describes it as a solid tool giving reliable answers and appreciates that chat history is stored locally. The criticisms are practical rather than ideological: that downloading the model requires creating an account, which grates against the fully-open framing even though the account belongs to the hosting platform rather than to the model; that responses tend to be verbose and slow compared with commercial services; and that there is no image generation, which is true — Apertus is a language model, and 1.5 added the ability to read images, not to produce them. Anyone arriving expecting a general-purpose assistant with the polish of a funded consumer product will be disappointed, and that expectation gap is what the low-ish rating mostly reflects.

Where it fits

Apertus is the right choice when the provenance of the model matters as much as its output: public-sector deployments, regulated industries, research that needs to describe exactly what the model was trained on, and work in languages that commercial models handle poorly. It is also the right choice for anyone who wants to run inference entirely on their own hardware with no licence negotiation, since Apache 2.0 imposes no usage restrictions. It is the wrong choice if you want the strongest available reasoning at any cost, if you need image generation, or if you want a polished consumer product with a support contract. The catalogue lists the platforms honestly as online and self-hosted, with no desktop or mobile clients of its own, and features that are consequences of the licence rather than of engineering: no tracking, no registration for the chat, offline capability once you host it yourself. If your task is condensing documents rather than choosing a foundation model, a hosted summarisation service such as SumizAI is a shorter path than standing up 70 billion parameters of infrastructure.

Product page: apertus-ai.org

What SumizAI is

A note-taking application built around a conversation with an AI model. You ask a question, the answer streams back, and the answers worth keeping become Markdown notes — filed into a vault that is a folder on your own disk.

A vault is a directory holding base.md, a generated table of contents up to six levels deep, and a notes/ folder with one .md file per note. Before a note is written it is checked against the ones already there, so the fourth note about the same idea gets merged instead of added. Links between notes are ordinary Markdown links to files that exist.

The model is never ours: you bring your own API key to one of seven providers — Anthropic, OpenAI, Gemini, Groq, OpenRouter, a local Ollama or your own server — and pay that provider directly. A question sends the table of contents plus at most five relevant notes within a 24,000-character budget, and the app shows you which five it used.

Every answer kept this way becomes one of these ai notes, ready to search again months later without hunting back through old chats.

Side by side

Where they differ

• Apertus is a place to have the conversation. SumizAI is not a chatbot — the conversation is the input, not the product. What comes out of it is a .md file with a title, a place in base.md and a duplicate check against the notes already in the vault. If you only want to talk to a model, you do not need SumizAI.
• AlternativeTo's description of Apertus does not mention an AI model. That is the whole starting point in SumizAI: you have a conversation with a model on your own API key, and the answers worth keeping become notes without you filing them.
• Apertus is open source (Apache-2.0), and SumizAI is not. If reading the source is what decides it for you, that is a real argument for Apertus. The guarantee SumizAI offers instead is structural rather than legal: the notes are Markdown files in a folder you chose, so they open in any editor whether or not the application is running.
• AlternativeTo's description of Apertus does not mention Markdown, so check what format your notes end up in before you fill it up. SumizAI writes standard .md files to a folder you picked, and an unpaid account can still export all of them.
• Apertus can be self-hosted. SumizAI has nothing to host on the desktop — the vault is a folder on your machine — and if you want the model on your own hardware too, a local Ollama is one of the seven providers it speaks to.
• Apertus runs in the browser. SumizAI is a desktop and mobile application, and on the desktop the vault is a folder on your own disk rather than a document in someone's cloud.

Which one to pick