Key Findings
Apertus 1.5 introduces native multimodal capabilities to the 8B and 70B model weights, allowing the processing of images and audio alongside text. The context window has been increased fourfold to 262,144 tokens. Performance improvements are noted in instruction-following, tool use, and reasoning-heavy benchmarks such as AIME when using the optional thinking mode. Specifically, the 8B model outperforms similarly sized open-weight models on multilingual tasks and IF, while matching GPT-OSS-20B and outperforming Claude Haiku 4.5 and Google Translate on the Last Translation Benchmark.
Technical Details
The models utilize an early-fusion approach for multimodality, treating images and audio as discrete tokens within the same sequence as text. Images are processed via the Emu3.5 tokenizer (512×512 images become 1,024 tokens) and audio via WavTokenizer (40 tokens per second). The architecture remains unchanged, though the vocabulary is extended.
Pretraining involved adding 4 trillion tokens to the 8B model (totaling 19T) and 2 trillion tokens to the 70B model (totaling 17T) across four stages: Unpaired, Paired, Cooldown, and Long context. Post-training followed a five-stage pipeline: SFT (using 3.7M text samples), long-context SFT, RL with verifiable rewards (RLVR) on 350k prompts, DPO, and charter alignment (SDPO).
Background & Context
Apertus 1.5 is a continued pretraining of the first-generation models. The project maintains a commitment to open weights, open data, and open values, adhering to the Apertus Charter. Text data was refreshed using sources including FineWeb2-HQ+, FinePDFs, and Nemotron-CC, ensuring all data is publicly available and respects robots.txt opt-outs.
Significance & Outlook
The release positions Apertus 1.5 as a competitive open-weight alternative for spatial and document understanding, with multimodal performance comparable to Gemma 4 12B. By collaborating with the EvalEval Coalition and publishing results in the Every Eval Ever (EEE) schema, the project provides transparent and reproducible benchmarking for researchers and developers.
Source: https://huggingface.co/blog/swiss-ai/apertus-v15-release-notes
Get our weekly technology intelligence — free
Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.
Subscribe Free — Weekly Tech Intelligence
By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.
- Your email and selected fields are used only to deliver the newsletter.
- We never share your information with third parties.
- You can unsubscribe anytime via the link in each email.
See our Privacy Policy for details.
Takes about a minute · Unsubscribe anytime
