Swiss AI Opens Apertus 1.5 With 262K Context, Multimodal Weights and 4T Added Pre-Training Tokens

Switzerland's Swiss AI released Apertus 1.5 under Apache 2.0, continuing pre-training of its 8B model with 4T additional tokens (19T total) and the 70B with 2T added (17T total). The update adds native multimodal input across images and audio, an optional thinking mode, better tool use, and extends context to 262,144 tokens — a 4x jump from Apertus 1.0. Base, SFT, RLVR, DPO and SDPO intermediate checkpoints are all public on Hugging Face.