Apertus 1.5 Adds Multimodal Reasoning and 256K Context to Open Weights

Apertus 1.5 Brings Multimodal Reasoning to Open Weights

Swiss AI's Apertus 1.5 release upgrades the 8B and 70B parameter models with native image and audio understanding, a configurable thinking mode, and a context window expanded to 262,144 tokens. The Apache 2.0 licensed weights arrived on Hugging Face in July 2026, continuing the project's commitment to fully open development — open weights, open data, and transparent training methodology.

Early-Fusion Multimodality Without Architecture Changes

Instead of bolting on separate vision or audio encoders, Apertus 1.5 tokenizes images via Emu3.5 (512×512 → 1,024 tokens) and audio via WavTokenizer (40 tokens/second) directly into the text stream. This early-fusion approach keeps the transformer architecture unchanged, allowing multimodal training at text-only speed. The models comprehend visual and spoken input but generate only text, a design choice that simplifies deployment.

Training unfolded in four stages: unpaired modality exposure, paired alignment, a quality-focused cooldown with extra math and code, and progressive context extension from 8K to 262K tokens. Over half the 19 trillion (8B) and 17 trillion (70B) token corpus remains text-only, preserving core language capabilities while integrating new modalities from permissively licensed sources.

Why it matters for GPU / AI infrastructure: The 4× context expansion and multimodal token streams increase per-sequence memory footprint, favoring GPUs with high VRAM or multi-GPU setups for fine-tuning and inference. Open weights let teams optimize quantization, distillation, or custom kernels on their own hardware — whether on-premises clusters or cloud instances like those from AiGpu — without vendor lock-in.