Swiss AI's Apertus 1.5 open model adds image input and 262K context

The Swiss AI Initiative, a collaboration led by EPFL, ETH Zurich and the Swiss National Supercomputing Centre (CSCS), has released Apertus 1.5, an update to its fully open language model family that now accepts images and audio alongside text. The release ships in 8-billion and 70-billion-parameter sizes under the Apache 2.0 licence, and extends the context window to 262,144 tokens, four times that of Apertus 1.0.

The models are not brand new: Evoya, a Swiss AI platform that tracks the project, dates the release to 24 July 2026, and EPFL’s School of Computer and Communication Sciences announced it on LinkedIn on 28 July. What has changed since then is the evidence. The team’s technical report, dated 11 October 2026, now documents how the models were built and how they score. This article sets out what that report and the coverage confirm about the release date, specifications and benchmarks.

What Apertus 1.5 adds over Apertus 1.0

Apertus 1.5 is not a new architecture. The technical report describes it as continued pretraining of the Apertus 1.0 models, on 4 trillion additional tokens for the 8B model and 2 trillion for the 70B. AI Weekly puts the totals at 19 trillion tokens for the 8B and 17 trillion for the 70B. Evoya lists the original 1.0 training run at 15 trillion tokens.

The main changes, as the sources describe them:

• Image and audio input. The report says continued pretraining “adds image and audio understanding”. Evoya describes audio input as an experimental feature.
• Text-only output. Multimodality enters through the tokenizer (the component that turns input into tokens), while the language-model head, the part that produces output, “remains restricted to text”. The model can read an image, but it cannot generate one.
• Fourfold context. The context window, the amount of input the model can consider at once, grows to 262,144 tokens, up from 65,536 tokens in version 1.0.
• Thinking as a switch. Step-by-step reasoning is “a flag of the chat template rather than a separate model”, so developers toggle it per request instead of loading a second model.
• Better tool use. EPFL cites stronger reasoning, improved instruction following and enhanced tool use, which Evoya frames as the foundation for AI agents.

After pretraining, the team applied supervised finetuning, long-context finetuning, reinforcement learning with verifiable rewards (RLVR, training against answers that can be checked automatically) and preference alignment. Evoya also reports 16 compact variants under the name Apertus Mini, aimed at weaker hardware such as laptops.

Spec	Apertus 1.5	Source basis
Sizes	8B and 70B parameters	Evoya fact sheet
Licence	Apache 2.0	AI Weekly
Inputs	Text, images, audio	Technical report
Output	Text only	Technical report
Context window	262,144 tokens	Technical report
Added training tokens	4T (8B), 2T (70B)	Technical report
Languages	More than 1,000	Evoya fact sheet

Release date and who shipped it

The sources agree on the July timing but differ slightly on attribution. Evoya says CSCS released Apertus 1.5 on 24 July 2026. EPFL’s post credits researchers from EPFL, ETH Zurich and CSCS jointly. The technical report names the project as “a collaborative effort led by ETH Zürich, EPFL, and CSCS, with contributors from many further institutions”, and lists Antoine Bosselut, Martin Jaggi and Imanol Schlag as leads.

The report itself appeared on arXiv on 11 October 2026, and AI Weekly covered the release on 15 October. Evoya’s page, which states its figures are as of October 2026, still says “no benchmarks yet” and that Public AI had announced a technical report for the coming weeks. That page appears to predate the report’s publication, so readers relying on it will miss the newer numbers.

What “fully open” means here

Apertus positions itself against models that publish weights but nothing else. The report defines full openness across five axes: open weights, open data, open code, open recipes and open values. For 1.5, the team says it releases “the model weights, data preparation and training recipes, and evaluation suites under permissive licenses”, along with the Apertus charter that guides its alignment stages.

On data, the report states the team trains only on openly available data, respects crawler opt-outs retroactively and excludes restrictively licensed material, as it did for 1.0. AI Weekly reports that intermediate checkpoints from the Base, SFT, RLVR, DPO and SDPO training stages are all public on Hugging Face, which lets researchers study or fork the model at any point in its training.

The Apache 2.0 licence allows commercial use, and Evoya says licence and openness terms are unchanged from 1.0.

Benchmarks: what the technical report claims

All benchmark claims for version 1.5 in the sources come from the Apertus team’s own evaluation. No independent leaderboard results appear. In the technical report:

• The released models “improve over their predecessors on nearly every benchmark”.
• Among fully open models of its class, the 8B “leads on knowledge, instruction following, and cultural benchmarks”, and its instruction following “surpasses all compared models, including open-weight ones”.
• Turning on thinking “lifts mathematical reasoning by up to 20 points on the 8B model”.
• On image and audio, the models are “competitive with fully open multimodal models”, strongest on document understanding and on long-form and lower-resource speech recognition.
• The report concedes that specialized vision and audio models remain ahead on perception and on transcription of English and other high-resource languages.

Two caveats matter. First, the headline comparative claims in the abstract are about the 8B model; the excerpt does not give a matching ranking statement for the 70B. Second, “fully open models” is a narrow peer group. The report itself acknowledges that fully open models “still trail their open-weight peers in capabilities”.

The 1.0 baseline

For context, Evoya reproduces the 1.0 base-model results from the Hugging Face model card: averaged over six tests (ARC, HellaSwag, WinoGrande, XNLI, XCOPA, PIQA), Apertus 70B scored 67.5, against 67.3 for Llama 3.1 70B and 69.8 for Qwen2.5 72B. Evoya’s reading is that Apertus is roughly on par with Llama 3.1 and slightly behind Qwen2.5, and that it was not compared directly with the largest commercial models. Those numbers predate 1.5 and should not be read as 1.5 results.

How to access Apertus 1.5 70B

The sources describe several routes:

• Download the weights. The weights are on Hugging Face, including the 1.0 model under swiss-ai/Apertus-70B-Instruct-2509 and a collection for version 1.5. The 1.0 model’s default system prompt states a knowledge cutoff of April 2024; the sources do not give a cutoff for 1.5.
• Hosted APIs. Evoya reports that Infomaniak offers Apertus 1.5 70B via API, and that Swisscom has offered Apertus to business customers on its Swiss AI Platform since September 2025.
• Third-party gateways. apertus.ai, which says it is an independent company not affiliated with the Swiss AI Initiative, lists Apertus 1.5 70B behind an OpenAI-compatible endpoint.
• Browser chat. The non-profit Public AI offers a chat with the 8B model without sign-up and the 70B model with thinking mode after login, according to Evoya, which says it is not meant for confidential data.

One discrepancy is worth checking before building on a hosted version. apertus.ai lists the 70B model with a 65K-token context, far below the 262,144 tokens in the technical report. Hosted endpoints may cap context below what the weights support, so the advertised window of any given provider should be confirmed.

What it means for builders

For teams that need a model they can audit end to end, Apertus 1.5 adds the inputs that earlier versions lacked. Document understanding is where the report says the multimodal side is strongest, which fits workloads such as reading scanned contracts or forms. Evoya points out that long contracts or entire dossiers can now fit in a single request.

The limits are clear in the sources. The model outputs text only. Specialized vision and speech models still lead on perception and English transcription. Evoya cautions that large commercial models remain clearly stronger for complex tasks, long texts and programming, and that Apertus stands out for transparency and control rather than top performance, though that assessment was written around the 1.0 benchmarks.

Self-hosting the 70B is a server-class job. Evoya’s own rough estimate is around 140 GB of GPU memory at full precision. The 8B model, quantised (compressed to lower numeric precision), runs on a laptop, and Evoya cites Liip as saying a graphics card with 8 GB of memory is enough.

A commenter on EPFL’s announcement raised the governance angle: in regulated sectors, trust “also depends on governance, traceability, evidence and auditability at every inference”. Open training data helps with traceability, but it does not by itself supply runtime audit controls.

What to watch next

• Independent evaluations. All 1.5 benchmark claims so far are the team’s own. Third-party results, especially for the 70B on vision and long-context tasks, would test the report’s claims.
• Hosted context limits. Whether API providers expose the full 262,144-token window or a reduced one.
• Audio maturity. Audio input is described as experimental; how it develops, given the report’s strong results on lower-resource speech recognition, is worth tracking.
• Apertus Mini. How the 16 compact variants Evoya describes perform on-device.