The Swiss AI Initiative has released detailed October 15 notes and a technical report for Apertus 1.5, its 8B- and 70B-parameter multilingual model family. This is a documentation and reproducibility milestone: the team explicitly dates the underlying Apache-2.0 weights on Hugging Face to July 24, 2026. Describing October as the first model launch would be wrong. The new documentation clarifies how text-only Apertus 1.0 was extended to handle images and audio without abandoning the underlying decoder architecture.
The project uses discrete multimodal tokens in a single sequence. Swiss AI says images are encoded with Emu3.5 and audio with WavTokenizer; its release notes describe a 512-by-512 image as 1,024 tokens and audio at 40 tokens per second. Both 8B and 70B models accept text, image and audio inputs and generate text, not pictures or speech. Audio support is identified as experimental, so application-level quality and language coverage require their own checks.
Apertus 1.5 extends the context window to 262,144 tokens, four times that of the previous generation, and adds an optional thinking mode. The release notes trace continued pretraining with four trillion additional tokens for 8B and two trillion for 70B, plus staged long-context training. Swiss AI also links data documentation, intermediate checkpoints, training recipes and structured evaluation records. That transparency allows specialists to investigate claims more deeply than a weight-only release, but does not by itself establish independent performance.
Access still has a practical caveat. The official model repository displays Apache-2.0 licensing while requiring a logged-in agreement to an acceptable-use policy and sharing of contact details before downloading files. Inference guides also note that thinking-mode and tool-call settings are not currently combined in their recommended vLLM configuration. Users should verify those constraints, memory requirements and exact checkpoint revisions before choosing Apertus for regulated or reproducible deployments.
What this changes for the evidence layer
Treat 15 October as the documentation milestone, not the first weight-publication date: official Apertus v1.5 weights were published July 24. Validate separate swiss-ai/Apertus-v1.5-8B and -70B checkpoint revisions, file access conditions, Apache-2.0 text and linked acceptable-use policy before any registry change. The current Hub gate means visible metadata alone does not prove anonymous download. Publisher evaluation claims are not our measurements.
Swiss AI Initiative official Apertus 1.5 technical release notes
This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.
Open source article ↗
