Switzerland just put a number on how far “open” can go in large language models. Apertus 1.5, the second release from the Swiss AI Initiative, ships with 262,144 tokens of context, native image and audio understanding, and a training pipeline that is published down to the intermediate checkpoint. Built by EPFL, ETH Zurich, and the Swiss National Supercomputing Centre (CSCS), the model lands at a moment when governments and enterprises outside the United States and China are asking a blunt question: can we run frontier-adjacent AI without handing our data to a foreign cloud? Apertus 1.5 is Switzerland’s answer, and the technical paper behind it, posted to arXiv in October 2026, gives the clearest picture yet of what that answer actually contains.
What Is Apertus 1.5?
Apertus 1.5 is a family of two open-weight language models, sized at 8 billion and 70 billion parameters, released by the Swiss AI Initiative as a continued-pretraining successor to the original Apertus 1.0. According to the release notes published on Hugging Face, the project describes itself as one of the world’s largest fully open language model efforts, built entirely by ETH Zurich, EPFL, and CSCS. Unlike most “open” model releases that hand over only the final weights, Apertus publishes training data sourcing, data-preparation methods, training recipes, evaluation suites, alignment principles, and intermediate checkpoints spanning base, supervised fine-tuning, RLVR, DPO, and SDPO stages.
That level of disclosure is the point. Apertus is less a product launch than a sovereignty statement: a working demonstration that a mid-sized European country, without the capital of a Microsoft, Google, or Alibaba, can train, document, and ship a model competitive with other open-weight systems. The timing matters too. Apertus 1.5 arrives in the same window as Meta’s Muse Spark Open and Nvidia’s Nemotron 3.5 Super VL, both released on October 15, 2026, turning this week into one of the busiest stretches for open-weight model drops this year.
Inside the Numbers: Parameters, Context Window and Training Data
The jump from Apertus 1.0 to 1.5 is substantial on paper. The 8B model received 4 trillion additional training tokens, bringing its total to roughly 19 trillion tokens, while the 70B model received 2 trillion additional tokens for a total near 17 trillion. The context window grew to 262,144 tokens, a fourfold increase over the original Apertus, according to the AI Weekly release summary. Refreshed text sources for the new training run include FineWeb2-HQ+, FinePDFs, and Nemotron-CC, continuing the project’s practice of naming its data sources rather than treating them as a trade secret.
| Spec | Apertus 1.0 | Apertus 1.5 |
|---|---|---|
| Model sizes | 8B, 70B | 8B, 70B |
| Context window | ~65,536 tokens (estimated) | 262,144 tokens |
| Additional training tokens | N/A | 8B: +4T tokens; 70B: +2T tokens |
| Total training volume | Baseline run | ~19T tokens (8B); ~17T tokens (70B) |
| Modality | Text only | Text, image, and audio input |
| Reasoning mode | Not available | Optional thinking mode added |
| License | Apache 2.0 | Apache 2.0 |
The language coverage claims are where the numbers get harder to pin down precisely. The Swiss AI Initiative’s own materials describe training on more than 1,000 languages, while a separate analysis cited by AI Weekly puts the figure above 1,800 languages in the corpus, including Romansh and Swiss-German dialects. One report estimates roughly 40 percent of the training data is in languages other than English. The discrepancy likely reflects different counting methods (languages officially supported versus languages merely present in the corpus), but either figure dwarfs what most commercial frontier labs disclose about their multilingual training mix.
From Apertus 1.0 to 1.5: What Changed
Beyond the raw token counts, Apertus 1.5 adds two capability categories that the original release lacked entirely: multimodal input and a reasoning mode. Both moves track with where the rest of the open-weight field has moved in 2026, but the Swiss team frames them specifically around the project’s reproducibility commitment rather than chasing benchmark headlines.
Multimodal Inputs: Images and Audio
Apertus 1.5 extends beyond text to understand images and audio natively, according to the ETH AI Center’s announcement of the release. That puts it in the same broad category as Nvidia’s newly released Nemotron 3.5 Super VL and Nemotron Bolt variants, which Nvidia shipped the same day with 120 billion and 74 billion parameters respectively, the smaller Bolt model designed to run inside a 24GB memory footprint. For Apertus, native audio support is notable because few open-weight projects publish audio-capable weights at all, let alone alongside full training documentation.
Thinking Mode and Expanded Tool Use
The 1.5 release also introduces an optional thinking mode, giving the model a more deliberate reasoning trace before it answers, plus dedicated evaluation of tool-calling performance. The technical paper posted to arXiv in October 2026, titled “Apertus 1.5: From Open Text Models to Open Multimodal Reasoners,” frames this as the project’s attempt to stay relevant for agentic workloads without abandoning its transparency mandate. That is a harder balancing act than it sounds: reasoning traces and tool-use fine-tuning are exactly the kind of proprietary technique most labs guard closely, and Apertus is publishing the recipe.
Why Apache 2.0 Licensing Matters
Apertus 1.5’s weights ship under the Apache License 2.0, a permissive license that allows commercial use, modification, and redistribution with minimal conditions. That distinguishes it from models that market themselves as “open” while attaching usage caps, field-of-use restrictions, or revenue thresholds that trigger a separate commercial license. The project goes further than the license text alone: publishing training data, data-preparation methods, training recipes, and intermediate checkpoints is a disclosure standard that exceeds what the license itself requires.
That distinction matters for the “is it really open source” debate that follows every major model release. The supplied documentation acknowledges the nuance directly: open weights under a permissive license are not identical to open source software, because training-data copyright and dataset licensing questions don’t map cleanly onto code licensing. Apertus’s answer is to publish everything it legally can, including data-owner opt-out handling, and let outside auditors judge the result rather than asserting an “open source” label and moving on.
The Data Sovereignty Argument
The strategic case for Apertus has little to do with topping a leaderboard. It is about control. The Swiss AI Initiative’s stated rationale is that giving domestic institutions a large, inspectable model reduces dependence on proprietary foreign providers while preserving control over data, infrastructure, model behavior, and regulatory compliance. For Swiss banks, hospitals, and government agencies bound by strict Swiss and EU data-protection rules, a model trained on a national supercomputer with a published data pipeline removes a category of legal and procurement risk that comes with routing sensitive queries through a US or Chinese API.
The project also emphasizes memorization controls and respecting data-owner opt-outs retroactively where possible, a detail aimed squarely at the copyright and privacy objections that have dogged nearly every large training run since 2023. None of this guarantees Apertus will be adopted at scale. But it reframes the pitch: instead of competing purely on benchmark score, Apertus is competing on the question of who gets to audit the system a country’s institutions run on.
Inside Switzerland’s AI Infrastructure
Apertus 1.5 was trained on the Alps supercomputer in Lugano, operated by CSCS, under a collaboration spanning EPFL, ETH Zurich, and CSCS collectively known as the Swiss AI Initiative, according to the official project site at apertus-ai.org. That compute base is smaller than the dedicated AI clusters run by hyperscalers, which is precisely why the project leans so heavily on efficiency choices: a 70B parameter ceiling rather than a trillion-parameter flagship, continued pretraining rather than a from-scratch run, and heavy reuse of permissively licensed public data rather than costly proprietary licensing deals.
It is a modest-infrastructure strategy by design. Where Meta’s Muse Spark Open needed a cluster of two Mac Studios just to run locally at its 816-billion-parameter scale, Apertus 1.5’s 8B and 70B models are built to run on far more accessible hardware, a deliberate trade-off that favors deployability within national institutions over raw scale.
How Apertus 1.5 Compares to Rival Open Models
October 2026 has turned into a genuine open-weight model glut. Within days of each other, Meta, Nvidia, and the Swiss AI Initiative all shipped new open or open-weight systems, each optimizing for a different constraint: Meta for massive total capacity with low active compute, Nvidia for a multimodal model line scaling from enterprise GPUs down to 4GB-memory edge devices, and Apertus for transparency and sovereign deployability.
| Model | Developer / Origin | Parameters | Key Differentiator |
|---|---|---|---|
| Apertus 1.5 | Swiss AI Initiative (EPFL, ETH Zurich, CSCS) | 8B / 70B | 262K context, full data/training transparency, Apache 2.0 |
| Meta Muse Spark Open | Meta | 816B (42B active, MoE) | 1M-token context, runs across two Mac Studios |
| Nvidia Nemotron 3.5 Super VL | Nvidia | 120B (12B active) | Multimodal, scales down to Nemotron Bolt E4B for ~4GB devices |
| Mistral Large 4 | Mistral AI (France) | 1.05 trillion | Commercial API pricing at $1.36 per million tokens |
| DeepSeek V4.1 Flash | DeepSeek (China) | Not disclosed in full | 1-million-token context window via API |
On raw parameter count, Apertus is the smallest model in this group by a wide margin. Mistral’s Large 4 alone is roughly fifteen times larger than Apertus 1.5’s 70B variant. That gap is the entire point: Apertus isn’t trying to out-scale commercial frontier labs. It’s trying to be the model a government auditor can fully inspect, which gets harder, not easier, as parameter counts climb into the trillions.
The Global Sovereign AI Race
Apertus is not Europe’s only sovereignty play, and Switzerland is not alone in treating model ownership as a national asset. France’s Mistral AI has built a commercially oriented European model company, with its Large 4 release pushing deeper into enterprise API pricing rather than full open-weight disclosure. China’s DeepSeek and Qwen families, including the recently released wave of Chinese open-source LLMs, have built significant international adoption on the back of aggressive open-weight releases, though their governance and data-sourcing practices sit under a different regulatory regime entirely. The UAE’s Falcon project, developed by the Technology Innovation Institute in Abu Dhabi, represents a Gulf-state version of the same national-AI-capability logic.
What separates Apertus from all three is documentation depth rather than market share. Mistral competes on commercial performance and enterprise contracts. DeepSeek and Qwen compete on raw adoption numbers and aggressive pricing. Apertus competes on the auditability of its entire training pipeline, a narrower pitch, but one aimed precisely at the public-sector and regulated-industry buyers who are the most wary of black-box foreign AI dependencies.
Benchmark Reality: Where Apertus Leads and Where It Lags
The technical paper’s own framing is candid: Apertus 1.5 improves over Apertus 1.0 on nearly every benchmark the team tested, and the 8B model reportedly leads comparable fully open models on knowledge, instruction-following, and cultural benchmarks, outperforming similarly sized open-weight competitors specifically on instruction-following and multilingual tasks. Coding and reasoning scores improved substantially from version 1.0, but the paper acknowledges the model still trails leading frontier open-weight systems overall.
The 70B model’s story is less flattering. It reportedly remains behind similarly sized open-weight competitors, particularly in programming, mathematics, and reasoning, according to the EurasiaReview analysis of the project one year on. That is a meaningful caveat for anyone hoping Apertus 1.5, which PublicAI’s project profile describes in detail, closes the gap with frontier systems outright. It does not. What it does is set a new bar for how much of a model’s internals a developer or regulator can actually inspect, at the cost of raw benchmark supremacy.
Industry and Expert Reactions
Reaction from the European tech and digital-sovereignty community has been notably more enthusiastic than the project’s own cautious official language. Carla Buenger, a technology executive and digital sovereignty advocate, wrote that “with the release of Apertus 1.5 we have a relevant back-up plan,” framing the model as insurance against future restrictions on access to US or Chinese frontier systems, in a LinkedIn post. She added that “it is a privilege in Switzerland to have the knowledge on how to build, train and deploy our own models.”
Thomas Taroni, a technology executive associated with PHOENIQS Cloud, described the broader Swiss stack built around the release, stating that “for the first time, every layer of the stack is Swiss open to the world, governed under Swiss law,” in a separate LinkedIn post. Technology commentator Ricky Jones went further, declaring that “the sovereignty gap is closed,” while noting in the same post that ETH Zurich’s own official wording on the release is considerably more cautious than that characterization.
Not every assessment leans toward hype. Marcel Salathé, a professor and digital health researcher, offered a more measured read, writing that “Apertus won’t dethrone GPT or Claude. But it offers complete transparency, supports rare languages, and respects data protection laws,” in a LinkedIn post framing AI sovereignty as a spectrum rather than a binary switch. That tension, between sovereignty advocates claiming the gap is closed and researchers insisting it is a matter of degree, is likely to define how Apertus 1.5 is received outside Switzerland’s own policy circles.
Market Impact: Enterprise and Government Adoption
The most immediate adopters will almost certainly be Swiss and broader European public-sector bodies, universities, and regulated industries such as banking and healthcare, where data residency and auditability requirements already push procurement toward models that can be run on-premises or within sovereign cloud environments. The Apache 2.0 license removes the legal friction that might otherwise block commercial deployment, and the project’s publication of training recipes gives compliance and security teams something concrete to review rather than a vendor’s marketing claims.
For enterprises weighing Apertus against commercial frontier options, the calculus is less about raw capability and more about risk tolerance. A company already comfortable sending data to commercial frontier APIs has little reason to switch to a smaller open-weight model. A Swiss cantonal government, a European hospital network, or a defense-adjacent contractor bound by strict data-sovereignty mandates has a much stronger reason to consider it, even at the cost of some benchmark performance. That narrower but highly motivated buyer base is where Apertus 1.5’s commercial future will likely be decided.
Historical Context: The Open-Weight Movement
Apertus 1.5 sits inside a broader arc that has reshaped the AI industry since 2023, when open-weight releases were still treated as a secondary track behind closed, API-only frontier models. That dynamic has flipped repeatedly through 2026: open or open-weight releases from Meta, Nvidia, Mistral, DeepSeek, and now the Swiss AI Initiative have each, in their own way, argued that closed access is not a permanent requirement for state-of-the-art or near-state-of-the-art capability. Even OpenAI’s own roadmap shifts, including the retirement of GPT-5.5 in favor of newer Codex and Sol-branded releases, have happened against a backdrop where open alternatives are increasingly credible enough to pressure closed-model pricing and licensing terms.
What makes Apertus distinct within that movement is not the openness of the weights themselves, which Meta, Mistral, and DeepSeek have all also offered in various forms, but the openness of the process that produced them. Few other projects, commercial or academic, publish intermediate training checkpoints and data-preparation recipes alongside the final model. That is a research-transparency standard closer to how peer-reviewed science is supposed to work than how AI product launches typically operate.
What Comes Next: Five Predictions
- Expect other European nations and the EU’s broader sovereignty push to point to Apertus as a template, though few countries have Switzerland’s combination of CSCS-grade supercomputing and EPFL/ETH Zurich research capacity to replicate it quickly.
- Swiss public-sector procurement for AI tools will increasingly reference Apertus 1.5 as a baseline option, particularly in banking, healthcare, and cantonal government pilots where data residency is a hard requirement.
- Benchmark gaps in coding, mathematics, and reasoning versus frontier open-weight systems will narrow only gradually; expect an Apertus 2.0 cycle focused specifically on closing that reasoning gap rather than adding scale.
- The “is it really open source” debate will intensify as more projects claim the open label; Apertus’s full-pipeline disclosure standard is likely to become the comparison point critics use to call out shallower open-weight releases.
- Expect continued friction between enthusiastic sovereignty-advocate framing (claims the “gap is closed”) and more cautious academic assessments describing sovereignty as incremental progress rather than parity with US and Chinese frontier labs.
Frequently Asked Questions
What is Apertus 1.5? Apertus 1.5 is an open-weight large language model family, released in 8-billion and 70-billion-parameter sizes, built by the Swiss AI Initiative, a collaboration between EPFL, ETH Zurich, and the Swiss National Supercomputing Centre.
How big is Apertus 1.5’s context window? Apertus 1.5 supports up to 262,144 tokens of context, described as a fourfold increase over the original Apertus 1.0 release.
Is Apertus 1.5 really open source? Its weights are released under the permissive Apache 2.0 license, and the project additionally publishes training data sourcing, data-preparation methods, training recipes, evaluation suites, and intermediate training checkpoints. That exceeds typical open-weight disclosure, though training-data copyright questions mean “open source” in the strict software sense remains debated for any large language model.
Does Apertus 1.5 support images and audio? Yes. The 1.5 release adds native understanding of images and audio alongside text, extending the original text-only Apertus 1.0 models.
How does Apertus 1.5 compare to Meta’s Muse Spark Open or Nvidia’s Nemotron 3.5? Apertus is far smaller in parameter count than Meta’s 816-billion-parameter Muse Spark Open or Nvidia’s 120-billion-parameter Nemotron 3.5 Super VL. Its differentiation is transparency and sovereign deployability rather than raw scale.
Who trained Apertus 1.5 and where? The model was trained on the Alps supercomputer in Lugano, operated by the Swiss National Supercomputing Centre, as part of the joint EPFL, ETH Zurich, and CSCS collaboration known as the Swiss AI Initiative.
Does Apertus 1.5 outperform models like Llama, Mistral, or Qwen? According to the project’s technical paper, the 8B model leads comparable fully open models on instruction-following and multilingual benchmarks, but the 70B model still trails similarly sized open-weight competitors in programming, mathematics, and reasoning.
Why does Apertus 1.5 matter for AI sovereignty? It gives Swiss and European institutions a large, fully inspectable model they can run without depending on proprietary US or Chinese providers, which matters for compliance with strict data-protection and residency requirements in regulated sectors like banking and healthcare.