Apertus 1.5 Launches with 262K Context, Multimodality, and Advanced Reasoning

Apertus 1.5 marks a significant evolution in the open-source large language model landscape, representing a massive technical leap over its predecessor. This release is not merely an incremental update but a comprehensive overhaul, focusing on scale, multimodal understanding, and advanced reasoning capabilities. The core of the update involves a substantial increase in training data, with the 8B parameter model undergoing pretraining on an additional 4 trillion tokens, and the 70B model receiving 2 trillion tokens. This brings the total training corpus to 19 trillion and 17 trillion tokens, respectively. The commitment to fully open values—open weights, open data, and transparent training details—reinforces Apertus's position as a leader in accessible, research-grade AI.

Technically, the model's architecture has been enhanced to handle vastly more complex inputs. The context window, which dictates how much information the model can remember and process in a single conversation, has been dramatically expanded to 262,144 tokens, representing a fourfold increase from Apertus 1.0. This massive context capacity allows the model to process entire books, lengthy legal documents, or extensive codebases without losing coherence or forgetting earlier details. Furthermore, the model introduces a specialized 'Thinking Mode.' This is a crucial feature that allows the model to internally reason about the input—to simulate a step-by-step thought process—before generating the final answer. This mechanism significantly boosts performance, particularly on complex reasoning problems that require logical deduction, moving the model beyond simple pattern matching.

One of the most notable technical advancements is the native multimodal understanding. Apertus 1.5 is no longer limited to text; it seamlessly accepts images, diagrams, and audio alongside written prompts. This capability is achieved through an 'early-fusion' method. Instead of using a separate, isolated vision encoder (which processes images independently and then feeds results to the text model), the images and audio are tokenized and introduced directly into the main sequence of text tokens. For instance, an image is tokenized using the Emu3.5 tokenizer, turning a 512x512 image into 1,024 discrete tokens, while audio is processed via WavTokenizer. This integration is highly efficient because it only extends the model's vocabulary, leaving the core architecture and training speed unchanged, allowing the model to understand rich sensory data while still generating only text output. The model's ability to process spoken language, while still experimental, points toward future advancements in real-time conversational AI.

To achieve this state-of-the-art performance, the development process was meticulously structured across multiple stages. The pre-training phase itself was divided into four distinct stages: 1) Unpaired data, where the model learns to associate standalone images and audio with text; 2) Paired data, which aligns specific image-caption or audio-transcript pairs with language; 3) Cooldown, focusing on higher-quality, specialized data, particularly in mathematics and coding; and 4) Long context training, which systematically scaled the context window up to 262k tokens. Crucially, over half of the total training tokens remained text-only, ensuring that the robust language abilities established in Apertus 1.0 were preserved and enhanced.

Following pre-training, the model underwent five rigorous post-training stages. These stages refine the raw, massive knowledge into a usable, safe, and highly capable product. The process begins with Supervised Fine-Tuning (SFT) on 3.7 million text samples, supplemented by image and speech data. This is followed by Long-Context SFT, which is vital for restoring the long-context ability that can sometimes be lost during shorter, focused fine-tuning runs. Next is Reinforcement Learning with Verifiable Rewards (RLVR), which uses verifiable rewards across math, code, and instruction-following tasks, mixing thinking-on and thinking-off modes. The model then undergoes Direct Preference Optimization (DPO), using judge-scored preference pairs to dramatically improve chat quality, tool use, and overall safety. Finally, the process concludes with Self-Distillation Preference Optimization (SDPO), which aligns the model with the core Apertus Charter, ensuring transparency and adherence to defined ethical and operational principles. The combination of these five specialized post-training stages is what transforms a massive dataset into a reliable, predictable, and highly functional AI assistant.

Beyond the core model improvements, Apertus 1.5 significantly expands its evaluation suite. The evaluation process now includes dedicated multimodal evaluation pipelines, merging major frameworks like lmms-eval and VLMEvalKit. This comprehensive testing ensures that the model's performance is measured not just on traditional text benchmarks, but also on its ability to interpret complex visual and auditory information. Performance benchmarks show that while Apertus 1.5 may still trail the absolute frontier models, it demonstrates highly competitive performance. Specifically, the 8B version shows superior instruction-following and multilingual capabilities compared to similarly sized open-weight models. Its multimodal performance is noted to be competitive with models like Gemma 4 12B, and its ability to understand spatial relationships and interpret complex documents and charts is highly regarded within the open-source community. Furthermore, the model has shown strong, industry-relevant performance, matching leading open-weight models on complex tasks like Swiss Legal document analysis, and achieving machine translation capabilities comparable to GPT-OSS-20B. This deep dive into capabilities and transparent evaluation solidifies Apertus 1.5 as a benchmark for open, powerful, and versatile AI systems.