Physix Frontier badgePhysix Frontier
PublishChinese | EN

Physix Frontier · News Briefing Card (HuggingFace · Oct 15, 2026)

Swiss AI Releases Apertus 1.5 with Multimodal Input and Thinking Mode

KEY FACTS

  • The Swiss AI initiative released Apertus 1.5, a continued pretraining version of the original model.
  • The 8B model adds 4 trillion training tokens, while the 70B model adds 2 trillion.
  • The models natively support image and audio input, but output text only.
  • The context window expands to 262144 tokens, a 4x increase over version 1.0.
  • A new thinking mode lets the model reason before answering to improve reasoning performance.

KEY DATA

19T8B model total training tokens
17T70B model total training tokens
262144 tokensContext window
4xContext window increase

PHYSIX OBSERVATION

Apertus 1.5 packs multimodality and thinking mode into open weights and discloses all training details, a genuine boon for researchers who want to build their own compliant models. But images and audio go in only, never out, showing it remains positioned as an understanding-focused base rather than a generative multimodal model, and it still has ground to make up on leading open-source models.

Source: HuggingFace report

Swiss AI Releases Apertus 1.5 with Multimodal Input and Thinking Mode | Physix Frontier Briefing