Swiss AI publishes Apertus 1.5 release notes: open 8B and 70B models with image/audio input, thinking mode and 262k context under Apache 2.0

The Swiss AI initiative (swiss-ai) published detailed release notes on 15 October 2026 for Apertus 1.5, an open model family in 8B and 70B sizes. The weights were already on Hugging Face under Apache 2.0 from 24 July, about 11 weeks earlier. Compared with Apertus 1.0, the models had further pretraining (4T extra tokens for the 8B, 2T for the 70B), a 262,144-token context window (4x larger), an optional thinking mode, and image and audio input (text-only output). A technical report, a text SFT mix, a preference pool, training pairs and evaluation results were released alongside the weights.

Researchers and organisations that need to inspect, audit or reproduce a model's training can use the open data and evaluation records, which few comparable releases offer. Teams in regulated sectors, especially in the EU, may find the documentation useful as a starting point for transparency reporting. For most practitioners looking for the best open-weight model, the impact is limited, because the authors admit it trails frontier open-weight models. The release gives a credible option where openness matters more than raw performance.

This is not a new launch: the weights are 11 weeks old, and these notes mostly repackage that release. The regulatory angle is the authors' framing. Supporting EU AI Act reporting is not the same as demonstrating compliance, and no regulator has endorsed it. "Fully open" and the multimodal features also say nothing about quality until independent tests are run, and the notes list no direct comparison links in the extracted material.

1. 01Independent benchmark results for the 8B and 70B models against Gemma 4 12B and GPT-OSS-20B, especially on long-context (beyond 100k tokens) and audio and image tasks, within the next 4 to 8 weeks.
2. 02Hugging Face download counts, fine-tunes and derivative models built on Apertus 1.5, compared with Apertus 1.0 at a similar age, as a measure of real adoption.
3. 03Whether any EU body, public institution or regulated enterprise cites or deploys Apertus 1.5 for AI Act or GPAI Code of Practice documentation, and whether the training-data claims hold up under outside scrutiny.