Apertus 1.5 Moves Open Text Models into Multimodal Reasoning

Introduction

Apertus 1.5 is presented as a step from open text models toward fully open multimodal reasoners. Developed by the Swiss AI Initiative, the release is accompanied by a 134-page technical report covering the data mixture, training setup, and ablations. Its broader argument is that models built by public institutions, trained on documented data, and released openly can serve as an alternative to dependence on a small number of commercial providers.

What the release changes

• A continuation of Apertus 1.0. Rather than starting from an entirely new language model, the team continued pretraining the previous models. The 8B version uses 4 trillion additional tokens, while the 70B version uses 2 trillion. The update adds image and audio understanding and expands the context window fourfold to 262,144 tokens.
• A lightweight multimodal design. Multimodality enters through the tokenizer. The core architecture remains unchanged and the language-model head is still restricted to text generation. In practical terms, the models can interpret images and audio and respond in text, but they are not presented as systems for generating images or speech.
• Reasoning controlled by the chat template. “Thinking” is enabled as a flag in the chat template rather than through a separate checkpoint. The report says this setting improves mathematical reasoning by up to 20 points on the 8B model. That number should be read in the context of the stated evaluations, not as a universal gain across every task.
• A multi-stage post-training pipeline. The released models go through supervised fine-tuning, long-context fine-tuning, reinforcement learning with verifiable rewards, and preference alignment. Together, these stages target instruction following, extended inputs, checkable solutions, and more useful conversational behavior.
• Strong but specialized multimodal results. Apertus 1.5 is competitive with fully open multimodal models on image and audio understanding. Its clearest advantages are reported in document understanding, long-form inputs, and lower-resource speech recognition. Specialized vision and audio systems still lead on fine-grained perception and transcription.

Why it matters

The release is significant because openness is treated as more than a downloadable set of weights. By documenting the training mixture, setup, and ablations, the report gives researchers material for replication, auditing, and further training. That level of disclosure is especially relevant to public-sector, educational, and research users who need to understand where a model comes from and how its behavior was produced.

The technical design also offers a pragmatic expansion path. Keeping the language-model body unchanged allows the team to reuse existing text capabilities while adding new modalities through the tokenizer and continued training. This may reduce architectural fragmentation and make long-context, multimodal, and reasoning work easier to study within one family.

Still, “open multimodal” does not mean “best at everything.” Apertus 1.5 is primarily a text-generating system with multimodal understanding. Users working on visual localization, detailed perception, high-accuracy transcription, or real-time applications should compare it with specialized models. Benchmark improvements are also not a substitute for checking data documentation, licensing, bias, and reliability in real deployments. The release’s larger message is that open models are moving from proving that training is possible to making multimodal expansion transparent and reproducible.

Source: Hugging Face Daily Papers