Anyone thinking open AI stopped at public weights needs to reconsider. Swiss-AI has released Apertus 1.5, a family of 8B and 70B parameter LLMs that redefines transparency: not only are the checkpoints available, but all training data, training procedures, and post-training recipes are public. In an industry where many “open” models merely share weights without details on data provenance, Apertus 1.5 raises the bar with a philosophy of full reproducibility.

The models, based on a decoder-only transformer architecture with xIELU activation and the AdEMAMix optimizer, result from continued pretraining starting from Apertus 1.0. The training mix added 4 trillion tokens for the 8B version and 2 trillion for the 70B, enriching the model with multimodal capabilities that now accept images, audio, and text as input and generate text. Simultaneously, the context window extends to 262,144 tokens, a fourfold increase over the previous release, allowing full processing of lengthy documents or conversations without chunking strategies.

A distinctive element is the thinking mode, which can be activated to improve performance on complex tasks. This is complemented by refined post-training for tool use and improved instruction following. It’s not just a technical feature: letting the model “reason” before responding opens up on-premise decision scenarios where reasoning quality is critical – from diagnostic support in healthcare to contract analysis in legal contexts, all cases where data cannot leave the enterprise perimeter.

The decision to publish everything – weights, data, training pipelines – has immediate consequences for those evaluating LLM self-hosting. A fully open model eliminates lock-in risk with vendors who do not disclose how they trained their models or what data they used. For organizations with regulated workloads (GDPR, public sector, finance) it means being able to conduct deep audits and fine-tuning customization with the assurance of not inheriting biases or data conflicting with internal policies. Swiss-AI has also implemented retroactive opt-out mechanisms and memorization prevention, an additional safeguard against legal risks related to accidental use of sensitive data.

From an infrastructure standpoint, the 70B version will require adequate hardware: several GPUs with ample VRAM for full-precision inference, though appropriate quantization can make it fit within more modest setups. The 8B model is already within reach of a single enterprise-grade GPU or even advanced workstations, enabling on-premise deployment for smaller teams as well.

The arrival of Apertus 1.5 also signals competitive pressure on other open-source players. It proves that native multimodality, long contexts, and reasoning capabilities are achievable without sacrificing radical transparency. The forthcoming technical report, with benchmarks and intermediate checkpoints, will provide further evidence to assess whether production performance matches expectations. Meanwhile, the ball is in the court of IT and legal teams: to verify that the promised transparency withstands a thorough audit and that the models, once deployed, meet expectations at scale.