A Community of 200 Researchers Just Gave Two Free AI Models Eyes and Ears
Project Apertus has released version 1.5 of its open, free-to-download AI models. They can now understand images and audio, think through hard problems, and read documents four times longer than before.
Key points
- Project Apertus released Apertus 1.5 on July 24, 2026, as free, open-weight models in 8-billion and 70-billion parameter sizes on Hugging Face.
- The 8B model trained on 19 trillion tokens total, with 4 trillion new tokens added for this version; the 70B model added 2 trillion, reaching 17 trillion total.
- Apertus 1.5's context window grew fourfold to 262,144 tokens, roughly the length of three average novels.
- On the Last Translation Benchmark, the 8B model matches GPT-OSS-20B and outperforms Claude Haiku 4.5 and Google Translate.
- All weights, training data sources, evaluation results, and post-training recipes are published under an Apache 2.0 licence, meaning anyone can reuse or build on them commercially.
A volunteer research collaboration called Project Apertus has updated its family of free AI language models to understand images and audio for the first time. Apertus 1.5 went live on Hugging Face on July 24, 2026, free to download with no usage restrictions.
Version 1.0 was text-only. Version 1.5 accepts photos, diagrams and spoken language alongside written text, though audio understanding is labelled experimental. Outputs are still text only.
How does it understand images and audio?
Rather than bolting on a separate image-reading module, the team used a technique called early fusion: images and audio are converted into the same numbered units, called tokens, that the model already uses for words. A 512 by 512 pixel image becomes 1,024 tokens using the Emu3.5 tokenizer, developed by a team at BAAI. Audio is compressed to 40 tokens per second using WavTokenizer. Because the underlying model architecture didn't change, training ran at the same speed as a text-only run.
Context window growth matters in practice. Doctors uploading long patient records, or lawyers feeding in full contracts, no longer hit a hard wall mid-document; the window expanded fourfold to 262,144 tokens.
A switchable thinking mode lets the model reason step by step before answering. On AIME, a competition-level mathematics benchmark, thinking mode improved scores noticeably. It's the same pattern we've tracked across the open-model wave: AI2Day reported on Apple researchers building a model that thinks in text, images and sound at once just the day before this release.
What makes Apertus different from other open models?
Total transparency, applied consistently. The team publishes not just the model weights but every data source, every evaluation score, and post-training recipes in a machine-readable format called the EEE schema, backed by the EvalEval Coalition. That schema covers 56 of the 60 evaluation-reporting elements required under Article 55 of the EU AI Act and the accompanying code of practice, which matters as European regulators begin asking model developers to prove their claims.
The paper is at arXiv:2610.15542.
Apertus 1.5 still trails leading open-weight frontier models overall. Where it genuinely stands out: translation, where the 8B model matches a model two and a half times its size (GPT-OSS-20B) and beats Claude Haiku 4.5 and Google Translate on the Last Translation Benchmark; Swiss legal tasks, where it competes with the best open-weight options; instruction-following across multiple languages, where it outperforms similarly sized peers.
| Benchmark | Apertus 1.5 8B result |
|---|---|
| Last Translation Benchmark | Matches GPT-OSS-20B, beats Claude Haiku 4.5 |
| Swiss Legal tasks | On par with leading open-weight models |
| Instruction-following (multilingual) | Beats similarly sized open-weight models |
| Multimodal spatial understanding | Competitive among open-weight models |
| AIME (math reasoning) | Improves with thinking mode on |
What's distinctive here isn't the benchmark table. It's the governance layer. Publishing evaluation cards that regulators can directly map to legal requirements isn't common practice yet, and if it spreads it could quietly raise the floor for how the whole field reports results. Worth watching.
Common questions
Can I run Apertus 1.5 on my own computer?
The 8B model is small enough to run on consumer hardware with a capable GPU, though the 262k-token context window will strain most home setups. AI2Day covered what to expect from local AI in detail earlier this month.
Is the training data safe to use in commercial products?
Yes. The team used only publicly available, permissively licensed data, respected robots.txt opt-outs retroactively, and published every source. An Apache 2.0 licence on the weights permits commercial use.
What is an open-weight model?
An open-weight model is one where the developer releases the actual numerical parameters of the AI, not just an interface to it. Researchers, companies and individuals can run the model themselves, inspect it, and build products on top of it without paying per query.
