Apertus 1.5: Switzerland's Fully Open Model Gains Vision, Audio and Reasoning

Apertus 1.5 is the Swiss AI Initiative's answer to a question many teams ask about open models: can a publicly funded group release not just weights but the data recipes, training code and evaluation suites for a model that sees images, hears audio, reasons step by step and reads long documents? The newly posted technical report, 134 pages by the Apertus project, says yes for the 8B and 70B sizes.

This post summarizes what the report claims, where the claims are the authors' own, and what to check before building on it. The model website lists the 1.5 release in July; the full technical report appeared on Hugging Face papers this week, which is why we cover the write-up now.

Quick answers

What changed from Apertus 1.0

Apertus 1.0, which we covered when it launched as a fully open foundation model for sovereign AI, was a text-only family. Version 1.5 keeps the models and continues their pretraining: 4 trillion additional tokens for the 8B and 2 trillion for the 70B. The abstract says that stage adds image and audio understanding and extends the context window fourfold to 262,144 tokens.

Then comes a post-training stack: supervised finetuning, long-context finetuning, reinforcement learning with verifiable rewards, and preference alignment. That mirrors the recipe used by most reasoning-capable open models, but with every stage documented. The authors write that the report covers data mixtures, training setup and ablations "to facilitate the global open source development."

Multimodal without a vision tower

The most interesting engineering choice is how the model sees and hears. According to the abstract, "multimodality enters through the tokenizer, the language-model head remains restricted to text, and the architecture is unchanged." The Hugging Face summary from one of the authors calls this a vision tower-free fusion setup.

In plain terms: instead of attaching a large separate image encoder with its own parameters, images and audio are converted into tokens the existing transformer reads, while the model still only generates text. The upside is simplicity and an unchanged inference stack. The downside the authors acknowledge is performance: specialized vision and audio models remain ahead on perception and on transcription of English and other high-resource languages.

Where Apertus 1.5 is strongest, per the report, is document understanding and long-form and lower-resource speech recognition. For a family trained on more than a thousand languages, that fits the project's multilingual pitch.

Thinking as a flag

Apertus 1.5 does not ship a separate reasoning model. According to the report, thinking is a flag in the chat template, and enabling it lifts mathematical reasoning by up to 20 points on the 8B model. That design has a practical benefit: one set of weights serves both fast answers and deliberate reasoning, and you choose per request. The cost is the usual one for reasoning modes, more generated tokens and latency when the flag is on.

The benchmark claims, with caveats

The authors state that the released models improve over their predecessors on nearly every benchmark. Among fully open models of its class, they say the 8B leads on knowledge, instruction following and cultural benchmarks, and that its instruction following surpasses all compared models, including open-weight ones. On image and audio understanding, a first for the family, they describe the models as competitive with fully open multimodal models.

Two cautions. These are the authors' own comparisons, selected from the models they chose to compare against. And "fully open" is a narrower field than "open weights": models like Qwen, Llama or Mistral releases share weights but not data. Our own look at the broader open-weight race, such as Mistral Large 4 and its open-weights preview and Reflection AI's open-weight push, shows how different those categories are. Read the tables in the report before assuming Apertus replaces a larger open-weight model for your task.

Openness, data and values

The report is explicit about data governance. As in 1.0, the team trains only on openly available data, respects crawler opt-outs retroactively and excludes restrictively licensed material. It releases weights, data preparation and training recipes, and evaluation suites under permissive licenses, and says openness extends to values: the Apertus charter guiding the alignment stages is published alongside the models.

That matters for teams with compliance duties. A model whose training data is documented and opt-out-respecting is easier to defend in a procurement review, which is the argument behind much of the European sovereign-AI conversation we explain in what sovereign AI means. Compare it with other European efforts such as Aleph Alpha's Kolibri, and see our coverage of Ecosia dropping Mistral open-weight models for how European buyers are weighing model choices.

Who should try it

• Researchers who need a reproducible baseline: the data recipes and ablations let you re-run or modify training stages.
• Public-sector and regulated teams that need documented data provenance and a permissive licence.
• Multilingual builders, especially for lower-resource languages and long-form speech.
• Document-processing pipelines that benefit from the 262,144-token window and strong document understanding.

Who should not expect miracles: teams that need best-in-class English transcription, fine-grained visual perception, or frontier coding and agent performance. For those, the report itself points to specialized or larger models.

How to try it

The project hub is apertus-ai.org, which lists the 1.5 release, an inference-ecosystem guide dated September 17 and the ACL 2026 paper for the original technical report. Model cards and weights are under the Swiss AI organization on Hugging Face. A sensible first test on your own hardware:

1. Start with the 8B for cost; it is the size the report highlights for benchmark leadership among fully open peers.
2. Run your own evaluation set with thinking off and on, and record both quality and token usage.
3. Test long documents near the context limit; claimed windows often degrade before the maximum.
4. For audio, compare against a dedicated speech model on your languages before committing.

Open weights, open source, fully open: why the labels matter

Three terms get blurred in model announcements, and the Apertus report sits at the strictest end.

• Open weights means you can download and run the model. The training data and code may be private, and the licence may restrict use. Many of the best-known downloadable models fall here.
• Open source in the traditional sense would also include the code to train and evaluate the model under a recognised licence.
• Fully open is the claim Apertus makes: weights, data preparation and training recipes, evaluation suites and the alignment charter are all published, and the data itself is restricted to openly available sources with opt-outs honoured.

Why does the distinction change what you can do? With open weights alone you can fine-tune and deploy. With a fully open release you can also audit what went in, reproduce a stage of training, remove a data source and retrain, or measure contamination between training data and your benchmark. For regulated industries and academic work, that auditability is the product. For a startup that just needs a good model next week, it may be irrelevant, and a larger open-weight model will probably win on raw capability.

There is also a cost angle. The report describes continued pretraining of 4 trillion tokens on the 8B and 2 trillion on the 70B. That is large for a public institution but small next to the training budgets of frontier labs, which is exactly why the authors frame the result as proof that infrastructure built in the open, by public institutions, is viable at scale. The honest reading is that Apertus shows a path, not that it closes the gap with the best commercial systems.

Questions builders will ask

Can I fine-tune it? The weights and recipes are released under permissive licenses according to the authors, so fine-tuning is the intended use. Confirm the exact licence on each model card.

Does thinking mode change the API? Per the report it is a flag in the chat template, so the same checkpoint serves both modes. Expect your serving stack to need template support rather than a different model.

Will it replace my current open model? Only test can tell. Use the report's claim that the 8B leads fully open peers on instruction following as a reason to include it in your bake-off, not as a result you can assume on your data.

Is multimodal output supported? No. The language-model head is restricted to text, so the model reads images and audio but writes text.

What is verified and what is not

Why it matters

Most capable models today are either closed or open-weight with undisclosed data. Apertus is a rare example of a state-backed academic effort publishing the whole pipeline, and 1.5 shows it can keep pace on features: multimodality, long context and reasoning. Whether it matches the best open-weight models on raw quality is a separate question the community will test. But as a documented, reproducible base, it gives universities, governments and companies a foundation they can audit and modify, which is the practical meaning of "sovereign" here.

Details are accurate as of October 15, 2026, based on the authors' abstract and the project website; benchmark numbers should be read from the full report.

Related reading

• Apertus: the fully open foundation model for sovereign AI
• What is sovereign AI? Layers, models and trade-offs
• Aleph Alpha Kolibri: sovereign open-weight MoE
• Mistral Large 4 open-weights preview
• Reflection AI's open-weight model
• Ecosia drops Mistral open-weight models