The Swiss AI team published full release notes and a technical report for Apertus 1.5, the update to its fully open 8B and 70B models. The weights went up on Hugging Face under Apache 2.0.
What’s new
- Images and audio in, text out. Early fusion: images and audio become tokens in the same sequence as text, with no separate vision encoder. Audio is still experimental.
- Thinking mode you can switch on for reasoning problems.
- 262,144-token context, 4x Apertus 1.0.
- Better instruction-following and tool use.
- Continued pretraining on 4T more tokens for the 8B (19T total) and 2T for the 70B (17T).
As always with Apertus, it’s open weights, open data, open values and full training details.
My take
“Open weights” gets thrown around a lot. Apertus is the real thing: you can see what it was trained on. An 8B multimodal model with 262K context is a nice fit for local agent work.
