Evidence at a glance
An Extension, Not a New Foundation
The Swiss AI Initiative released Apertus 1.5’s 8B and 70B weights on July 24, 2026, under the Apache 2.0 license. The update is not aimed at one isolated task: it extends Apertus 1.0 with image and audio understanding, an optional thinking mode, longer context, better instruction following, and improved tool use. The project involves ETH Zurich, EPFL, and the Swiss National Supercomputing Centre CSCS, and presents the model as part of an effort to build open, transparent AI infrastructure that institutions can adopt themselves.
The important distinction is that this feature list does not describe a wholesale model redesign. The core architecture remains unchanged, with the new capabilities added through continued pretraining and several stages of post-training. That shifts the question from what a new architecture can do to how new input modalities and longer tasks can be connected to an existing model—and makes the training process part of the system’s value proposition.
Images and Audio Join the Same Sequence
Apertus 1.5 uses early fusion: images and audio are converted into discrete tokens by their respective tokenizers, then placed in the same sequence as text. Images use the Emu3.5 tokenizer and audio uses WavTokenizer, rather than feeding into a separate vision encoder. The project says that expanding the vocabulary leaves the core architecture unchanged and lets training run at text-only speed. The model still generates text only; it does not generate images or audio.
This design gives multimodal inputs a shared interface, but it also puts their costs on the same context budget. The release notes give a concrete scale: a 512-by-512 image becomes 1,024 tokens, while audio consumes about 40 tokens per second. Long documents, pictures, and recordings therefore compete for sequence space rather than receiving separate capacity. The project also describes speech understanding as experimental.
Long Context Is Not Free Multimodal Capacity
Apertus 1.5 has a context window of 262,144 tokens, four times the size of Apertus 1.0’s. That is useful for workflows that combine long documents with charts or audio, but it should not be read as a guarantee that the model can understand an equivalent volume of multimodal material with equal reliability. Images and audio already consume tokens, so a more varied input leaves less room for text and subsequent material.
A larger advertised limit is not enough to preserve long-context performance. The project says continued pretraining moved through unpaired text, images, and audio; paired image-caption and audio-transcript data; higher-quality math and code data; and a gradual expansion of the context window from 8k to 262k. Post-training added SFT at the full context length because shorter-context SFT can erode long-context ability. Deployment evaluations should therefore use realistic input lengths and modality mixes, not just check the maximum window listed on the model card.
Open Training, With Layers of Reproducibility
This is continued pretraining from Apertus 1.0, not a model trained from scratch. The 8B version adds 4 trillion training tokens, bringing its stated total to 19 trillion; the 70B version adds 2 trillion, for a total of 17 trillion. More than half of the training tokens remain text-only, which the project says helps preserve the previous generation’s language abilities. New text sources include FineWeb2-HQ+, FinePDFs, and Nemotron-CC. The release says new data must be publicly available and permissively licensed, and that robots.txt opt-outs are respected retroactively.
Post-training proceeds through SFT, long-context SFT, reinforcement learning with verifiable rewards (RLVR), DPO, and Charter alignment (SDPO). SFT uses 3.7 million text samples alongside image and speech data; RLVR uses about 350,000 prompts across math, code, instruction following, multilingual, and vision tasks. Tool-use training is restricted to trajectories verified by execution. The project also releases a data-source directory, its text SFT mix, and preference-training materials. But public source listings, selected datasets, and training details do not mean that every piece of training data is downloadable or that any institution can reproduce the run with the same resources.
Thinking Mode and Evaluation Need Context
Thinking mode is not a separate model. It is a setting switched through markers in the chat template, so support for thinking does not mean that every response automatically uses that path. Product integrations need to control the template and invocation explicitly. The project’s technical report says the 8B model improved by up to 20 points on math-reasoning evaluations, but the material available here does not provide the full benchmark, settings, or score changes. That figure should not be generalized to every reasoning task.
The project also says that Apertus 1.5 8B matched GPT-OSS-20B on the Last Translation Benchmark and outperformed Claude Haiku 4.5 and Google Translate. The provided material does not include the scores or testing conditions. For technical leads, this is a useful pointer for choosing an internal evaluation task, not a substitute for one. A more defensible conclusion is that Apertus 1.5 belongs on the shortlist when a team needs inspectable data sources, open weights, and image or audio input. The available evidence does not justify adopting it outright for workflows that depend on reliable speech understanding, multimodal generation, or demonstrated frontier performance.