The Release Notes for Apertus 1.5 Are Now Available

Apertus 1.5 is a continued pretraining of the first-generation open model family, adding 4 trillion tokens of text and multimodal training data to the 8B model and 2 trillion tokens to the 70B model, for totals of 19T and 17T respectively. The release keeps the project's full-open posture: open weights, open data, open values and complete training details. For anyone who followed Apertus 1.0, the headline is that the family has moved from text-only models to multimodal reasoners without changing its architecture.

What actually changed

The most visible addition is native multimodal understanding. Apertus 1.5 accepts images and audio alongside text, so documents, diagrams and photos can enter a conversation directly. Spoken language processing is included but the team describes it as experimental.

A thinking mode can be switched on so the model reasons about the input before answering, which the release says improves performance on reasoning problems. The context window grows to 262,144 tokens, a 4x increase over Apertus 1.0. Instruction-following is described as better than in 1.0, giving more predictable and accurate responses, and the models are trained for better tool integration so external tools and APIs can be used more effectively. The release also claims stronger adherence to the Apertus Charter, the document that sets out the model's values and principles.

Multimodality without a separate vision encoder

The design choice worth understanding is how images and audio get into the model. Apertus 1.5 uses an early-fusion approach: images and audio enter as discrete tokens in the same sequence as text tokens, rather than passing through a separate vision encoder. Images are tokenized with the Emu3.5 tokenizer, where a 512×512 image becomes 1,024 tokens, and audio with WavTokenizer at 40 tokens per second.

Because only the vocabulary is extended, the architecture stays unchanged and training runs at text-only speed. One limitation follows from this: the models understand images and audio but generate only text.

How the training was staged

Pretraining starts from Apertus 1.0 and runs in four stages. First comes an unpaired stage of text plus standalone images and audio, so the model learns the new modalities. A paired stage follows with image–caption and audio–transcript pairs to align them with language. A cooldown stage uses higher-quality data with more math and code, and a long-context stage grows the window step by step from 8k to 262k tokens.

More than half of the training tokens are text-only, which the team says preserves the language abilities of Apertus 1.0. Text data is refreshed with sources including FineWeb2-HQ+, FinePDFs and Nemotron-CC. All new data, multimodal included, meets the same bar as before: publicly available, permissively licensed, and with robots.txt opt-outs respected retroactively. Every source is listed in apertus-data-transparency.

Post-training runs in five stages: SFT, long-context SFT, RL, DPO and charter alignment (SDPO). SFT used 3.7M text samples plus image and speech data, with reasoning traces added only to prompts the model could not already solve without them, and tool-use data restricted to execution-verified trajectories. The text SFT mix is released on Hugging Face. Long-context SFT at the full 262k window restores long-context ability that short-context SFT erodes. RL with verifiable rewards (RLVR) ran on about 350k math, code, instruction-following, multilingual and vision prompts, mixing thinking on and off. DPO with judge-scored preference pairs targets chat quality, tool use and safety, and the preference pool and training pairs are released. SDPO aligns the model with the Apertus Charter through judge feedback and self-distillation.

Evaluation scope and where the models stand

The release significantly expands evaluation beyond the text benchmarks used for Apertus 1.0, adding multimodal evaluations for image and audio and thinking-mode evaluations. Multimodal testing runs through the MLLM-eval-suite pipeline, which merges the lmms-eval and VLMEvalKit frameworks. A series of text benchmarks was added too, along with extended benchmarking for translation and multilingual capabilities, safety and tool use, published in a post-training evaluation repository based on a fork of lm-evaluation-harness.

The team is direct about the gap to the frontier: Apertus 1.5 still trails frontier open-weight models. Within its size class, the 8B model outperforms similarly sized open-weight models on instruction-following and multilingual tasks, and thinking mode further improves results on reasoning-heavy benchmarks such as AIME. Multimodal performance is described as closer to Gemma 4 12B and other current open-weight models, competitive in spatial understanding among open-weight models and in document and chart understanding among open-source models.

External evaluations add two data points. On Swiss Legal tasks, Apertus 1.5 is described as on par with leading open-weight models. On the Last Translation Benchmark, the 8B model's machine translation matches GPT-OSS-20B and outperforms models such as Claude Haiku 4.5 and Google Translate.

Evaluation cards and the transparency argument

Apertus works with the EvalEval Coalition on benchmarking, publishing Apertus 1.5 results in the Every Eval Ever (EEE) schema. That format records not just scores but who ran the evaluation, how the model was accessed, the generation settings and how each metric should be interpreted — details routinely missing from leaderboards that often explain why similar-looking scores differ. Results appear in an Eval Card combining run data with benchmark and model metadata, and are also available through Hugging Face Community Evals. The same infrastructure is used by the UK AI Security Institute to publish its evaluations openly.

Because text evaluations run on the project's public fork of lm-evaluation-harness, which EEE supports with a dedicated converter, anyone can rerun them and compare numbers under identical settings. EvalEval has mapped Evaluation Cards to AI governance requirements including Article 55 of the EU AI Act and the GPAI Code of Practice, finding the EEE schema can represent 56 of the 60 evaluation-reporting elements those frameworks call for. This fits how Apertus already releases: alongside the models, it publishes a Public Summary of Training Content and, voluntarily, a Code of Practice document.

Reproducibility and what the ablations show

All recipes, ablations and data mixtures are documented in the Technical Report. Intermediate checkpoints are hosted in the Apertus 1.5 Hugging Face collection, covering base weights, SFT, RLVR and DPO, with curated training data mixes and custom datasets in the Apertus 1.5 datasets collection. Sharing them lets researchers investigate how each stage changes model behavior.

Two ablation findings stand out. Multimodal data mixed with text data during SFT helps improve across all modalities. And the benefits of multimodal reasoning traces survive subsequent text-only reinforcement learning: starting from an SFT checkpoint trained with multimodal traces, RL improves mathematics more than starting from a text-only checkpoint, and the advantage on reasoning tasks persists throughout training.

Availability and what comes next

Proton Lumo and Infomaniak Euria are two widely used apps that have recently integrated Apertus 1.5, linked in the Community Showcase. API access for developers has been announced by providers including Swisscom, Infomaniak and Phoeniqs, among others listed in the General Availability guide. Technical instructions for running the models are in the model cards on Hugging Face and in the Quickstart guide.

The project recently passed five million downloads in its Hugging Face analytics. At the AI+X Summit in Zurich, the team announced the roadmap for Apertus 2.0: a sparse/MoE architecture with roughly 600B total parameters, an extended training schedule, and agents, mathematics and coding as priorities.

For readers deciding whether to try Apertus 1.5, the practical takeaways are the 262,144-token window, image and audio input with text-only output, and a thinking mode that can be toggled. The open checkpoints and data mixes make it possible to inspect how each training stage shaped the model rather than taking the reported scores on faith.