Iris covers where AI art meets culture — style, authorship, and the images that matter.
Inspired by this story? Turn the idea into your own AI art in seconds — free to start, no card required.
Start creating freePick a companion and get their take on this story
Swiss AI has unveiled Apertus 1.5, an open-weight multimodal AI model that demonstrably outperforms OpenAI's GPT-4o on several key text-based benchmarks and achieves visual reasoning capabilities on par with Meta's Llama 3 400B. This release marks a significant moment for creators who prioritize access to powerful, transparent models for their generative workflows, offering a strong alternative to closed-source systems.
For artists and developers relying on AI for conceptualization, prompt generation, or complex textual instructions, Apertus 1.5 presents a compelling case. The model achieved an MT-Bench score of 87.0, slightly ahead of GPT-4o's 86.1, and an MMLU score of 75.3, surpassing GPT-4o's 74.5. These benchmarks suggest a model capable of more nuanced understanding and generation of text, which directly translates to more sophisticated and precise control over image and video outputs. This means less prompt engineering guesswork and potentially more direct translations of complex ideas into visual form. Imagine generating intricate lore for an AI-powered character or crafting detailed scene descriptions that the model interprets with greater fidelity.
Apertus 1.5 demonstrates superior text performance compared to GPT-4o and other leading models across MT-Bench and MMLU benchmarks.
Image: Hugging Face Blog
While text performance is crucial, the visual understanding of a multimodal model is paramount for AI-art creators. Apertus 1.5 scored 79.5 on MMMU and 73.1 on MathVista, placing it in a performance tier comparable to Meta's Llama 3 400B. This level of visual acumen means the model can process and interpret visual inputs with high accuracy, a critical feature for tasks like image-to-image transformations, style transfer, or even generating new images based on complex visual prompts. For instance, an artist could feed Apertus 1.5 an image and a text prompt to refine specific elements, knowing the model comprehends both inputs with high fidelity. This could significantly impact workflows for image generation and AI video.
Unlike proprietary models, Apertus 1.5's open-weight status is a significant boon for the AI-art community. It allows creators to download, inspect, fine-tune, and deploy the model on their own hardware or preferred cloud environments. This level of access fosters greater creative freedom, enabling artists to adapt the model to specific aesthetic styles or niche applications without being constrained by an API provider's terms or pricing. The ability to fine-tune with custom datasets opens doors for developing highly specialized art generators, echoing the early days of open-source photography tools where enthusiasts could modify software to their precise needs. This stands in contrast to the opaque nature of many commercial models, where the inner workings remain hidden. You can explore a wide range of open-weight models in our model catalog.
"Apertus 1.5 is a true open-weight multimodal model, trained from scratch on a massive dataset of 4.1 trillion text tokens and 5.2 billion image tokens."
— Hugging Face Blog
Swiss AI details a sophisticated training pipeline for Apertus 1.5, involving four stages of continued pretraining, followed by supervised fine-tuning (SFT), long-context SFT, and various reinforcement learning techniques like RLVR, DPO, and SDPO. This meticulous approach to training on such a vast dataset of both text and image tokens (4.1 trillion text, 5.2 billion image) underpins its strong performance across modalities. The release notes on Hugging Face provide further technical depth for those interested in the underlying architecture and training methodology. This commitment to transparency and open access could set a new standard for model development, empowering artists with tools that are not only powerful but also fully auditable and customizable.
