Apertus 1.5 Release Notes

What it claims

Apertus 1.5 continues pretraining from Apertus 1.0, bringing the 8B model to 19T total tokens (+4T) and the 70B model to 17T (+2T). It adds image and audio input; output is still text only, and audio is labelled experimental. It also adds a switchable thinking mode and extends context to 262,144 tokens, 4x Apertus 1.0. Weights, data, intermediate checkpoints and the SFT and DPO data mixes are released.

How it works

Images and audio become discrete tokens in the same sequence as text. The only change to the model is a larger vocabulary, so training runs at text-only speed.

flowchart LR
  I[Image] -->|Emu3.5 tokenizer| S[Shared token sequence]
  A[Audio] -->|WavTokenizer| S
  T[Text] --> S
  S --> M[Unchanged Apertus transformer]
  M --> O[Text output only]
  X[Separate vision encoder] -.not used.-> M

• Token cost: a 512×512 image takes 1,024 tokens (Emu3.5). Audio takes 40 tokens per second (WavTokenizer).
• Continued pretraining, four stages: 1. Unpaired text, images and audio. 2. Paired image–caption and audio–transcript data. 3. A cooldown on higher-quality data with more math and code. 4. Context extended stepwise from 8k to 262k.
• Text share: more than half of all tokens are text-only, to preserve 1.0's language ability. New text sources include FineWeb2-HQ+, FinePDFs and Nemotron-CC.
• Data policy: all data is publicly available and permissively licensed. robots.txt opt-outs are honoured retroactively, and sources are listed in apertus-data-transparency.
• Post-training:
• SFT on 3.7M text samples plus image and speech data. Reasoning traces are added only to prompts the model cannot already solve without them. Tool-use data is limited to execution-verified trajectories.
• Long-context SFT at the full 262k window, because short-context SFT erodes long-context ability.
• RLVR on about 350k prompts covering math, code, instruction following, multilingual and vision, mixing thinking on and off.
• DPO on judge-scored preference pairs.
• SDPO for charter alignment, using judge feedback and self-distillation.

Evidence

The post gives no scores, only qualitative comparisons. The numbers are in the Eval Card and the technical report.

• Overall: it still trails frontier open-weight models.
• 8B: beats similarly sized open-weight models on instruction following and multilingual tasks.
• Thinking mode: helps on reasoning-heavy benchmarks such as AIME.
• Multimodal: described as "closer to Gemma 4 12B". It is competitive in spatial understanding among open-weight models and in document/chart understanding among open-source models.
• Translation: on the Last Translation Benchmark, the 8B reportedly matches GPT-OSS-20B and beats Claude Haiku 4.5 and Google Translate.
• Swiss legal tasks: on par with leading open-weight models, according to external evaluators.
• Ablations from the report:
• Mixing multimodal data into SFT improves every modality.
• Starting text-only RL from an SFT checkpoint trained with multimodal reasoning traces improves math more than starting from a text-only checkpoint, and the advantage lasts through training.
• Reporting format: results are published in the Every Eval Ever (EEE) schema, which records evaluator, access method and generation settings. Text evals run on a public lm-evaluation-harness fork with an EEE converter, so you can rerun them under identical settings. Multimodal evals use MLLM-eval-suite, which merges lmms-eval and VLMEvalKit. EvalEval reports that EEE covers 56 of the 60 evaluation-reporting elements in EU AI Act Article 55 and the GPAI Code of Practice.

How to use it

• Checkpoints: the Apertus 1.5 Hugging Face collection has base, SFT, RLVR and DPO checkpoints. The datasets collection has the text SFT mix and the DPO preference pool and pairs.
• Running the models: instructions are in the model cards and the Quickstart guide. The page reproduces no commands.
• Hosted APIs: announced by Swisscom, Infomaniak and Phoeniqs, among others; the full list is in the General Availability guide.
• Deployments: Proton Lumo and Infomaniak Euria already use it.

Limits

• Audio understanding is experimental.
• The models cannot generate images or audio.
• Overall performance is below frontier open-weight models.
• The post gives no benchmark numbers or settings. Check the Eval Card or the report before relying on any comparison.
• The post does not state the weight licence.
• Budget for input length: images cost 1,024 tokens each at 512×512, and audio costs 2,400 tokens per minute.
• The announced Apertus 2.0 (MoE, about 600B total parameters, focused on agents, math and code) may change the picture if you are choosing a base model to build on.