ResearchAnalysis
Apertus 1.5 report: inside Switzerland's fully open model
The Swiss AI Initiative published detailed release notes for Apertus 1.5 on October 15, based on its technical report. The 8B and 70B models now understand images and audio, gain a 262k-token context and an optional thinking mode, and every training stage, dataset and checkpoint is public.
By: Mate · · 3 min read
Apertus 1.5, built by ETH Zurich, EPFL and CSCS, first shipped its weights on July 24 under Apache 2.0 (CSCS). What's new for researchers is the technical report (arXiv) and detailed release notes (Hugging Face).
This isn't just another open-weight model. The data, code and full training recipe are public too, which is rare.
What changed since 1.0
Version 1.5 continues pretraining from 1.0: 4 trillion new tokens for the 8B model and 2 trillion for the 70B, for 19T and 17T in total. New features include image and audio understanding (audio is still experimental), an optional thinking mode, a 262,144-token context window (four times 1.0's), and better instruction following and tool use.
Multimodality uses early fusion: images and audio enter as discrete tokens alongside text, with no separate vision encoder. A 512×512 image becomes 1,024 tokens; audio runs at 40 tokens per second. Only the vocabulary was extended, so the architecture is unchanged. The models understand images and audio but generate only text.
The research value: every step is downloadable
Post-training has five stages: SFT, long-context SFT, RLVR, DPO and, for the 8B model, charter alignment (SDPO). Intermediate checkpoints (base, SFT, RLVR, DPO) are published, most of them alongside the data mixes used.
Two findings stand out from the team's own ablations: mixing multimodal data into SFT improved every modality, and the benefit of multimodal reasoning traces survived later text-only RL, with maths improving more than when starting from a text-only checkpoint.
Why it matters: anyone trying to understand what each training stage does to a model can now compare checkpoints step by step. With closed models, that's impossible.
Performance, honestly
The team says plainly that Apertus still trails frontier open-weight models. By its own measurements, the 8B beats similarly sized open models on instruction following and multilingual tasks, and its multimodal performance is close to Gemma 4 12B. External evaluators report it on par with leading open models on Swiss legal tasks, and the 8B matches GPT-OSS-20B on one translation benchmark.
Results are also published in the EvalEval Coalition's Every Eval Ever schema, which records not just scores but who ran them and how. The team says the schema covers 56 of the 60 evaluation-reporting elements called for by Article 55 of the EU AI Act and the GPAI Code of Practice.
What's next
For Apertus 2.0, the team has announced a sparse/MoE architecture with about 600B parameters and a focus on agents, maths and coding. If EU compliance or reproducibility matters to you, Apertus is now one of the best-documented starting points, though it won't top the raw performance charts.
