Project Apertus announced Apertus 1.5, a continued-pretraining update to its 8B and 70B models that introduces native multimodal understanding, an optional thinking mode, and a fourfold increase in context length to 262,144 tokens. The team released the updated weights to Hugging Face on July 24, 2026 under an Apache 2.0 license and published a Technical Report with training details.
Key changes in the models
Apertus 1.5 accepts images and audio alongside text using an early-fusion approach: images and audio are tokenized and interleaved with text tokens rather than passed through a separate vision encoder. The release notes state that a 512×512 image becomes 1,024 tokens with the Emu3.5 tokenizer, and audio is tokenized at 40 tokens per second with WavTokenizer. The models understand these modalities but continue to generate text-only outputs.
The update also adds a selectable thinking mode intended to allow the model to reason about inputs before producing answers. Context capacity is extended to 262,144 tokens (a 4× increase versus Apertus 1.0). Project Apertus reports stronger instruction-following and improved tool integration in the new checkpoints.
Training, post-training and evaluation
The team says Apertus 1.5 is the result of continued pretraining: roughly 4 trillion tokens of additional text and multimodal data were added to the 8B model and about 2 trillion to the 70B model, yielding total pretraining volumes of 19T and 17T tokens, respectively. Pretraining proceeded in stages labeled unpaired, paired, cooldown and long-context. Post-training stages included SFT, long-context SFT, RL with verifiable rewards (RLVR), DPO, and charter alignment (SDPO).
Project Apertus expanded evaluations to cover multimodal and thinking-mode benchmarks, using the MLLM-eval-suite (merging lmms-eval and VLMEvalKit) and a public fork of lm-evaluation-harness. The team reports Apertus 1.5 8B outperforming similarly sized open-weight models on instruction following and multilingual tasks, competitive multimodal results near Gemma 4 12B on some metrics, and translation performance on the Last Translation Benchmark comparable to GPT-OSS-20B and ahead of models such as Claude Haiku 4.5 and Google Translate, according to their published runs.
Project Apertus also committed to reproducible reporting via Eval Cards and the Every Eval Ever (EEE) schema in collaboration with EvalEval, and released intermediate checkpoints, SFT mixes, preference pools and training recipes in accompanying repositories and model cards.
Availability: the Apertus-v1.5 weights and checkpoints are hosted on Hugging Face; the team lists Proton Lumo and Infomaniak Euria among early integrations and notes API access from providers including Swisscom, Infomaniak and Phoeniqs. The Technical Report and arXiv preprint (arXiv:2610.15542, October 11, 2026) contain full reproducibility details.
Original source: Hugging Face Blog