🔍 Read the full analysis: Apertus 1.5: New Features And Changes on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Apertus 1.5 was announced on July 24, 2026, with image and audio input, a context window expanded to 262,144 tokens, and reported gains in reasoning and instruction-following. The 8B and 70B models were released on Hugging Face under an Apache 2.0 license; speech understanding remains experimental, and real-world performance has yet to be established.
The Apertus project announced Apertus 1.5 on July 24, 2026, adding image and audio input and expanding the context window to 262,144 tokens in its 8B and 70B models. Released on Hugging Face under the Apache 2.0 license, the models give developers and researchers access to new multimodal capabilities in open-weight systems, a development that reflects broader questions about AI making intelligence less scarce, though the project describes speech understanding as experimental.
Apertus 1.5 builds on the earlier models through additional multimodal pretraining. The source says the 8B model received 4 trillion tokens of multimodal data and the 70B model received 2 trillion, bringing their total training amounts to 19 trillion and 17 trillion tokens, respectively. The project says images and audio can be processed alongside text as tokens in a shared sequence.
The release uses the Emu3.5 tokenizer for images and WavTokenizer for audio, according to the source. The models are intended to interpret material such as documents, diagrams, photos and spoken language. Audio capability is included, but speech remains experimental; the announcement does not establish how well it works across accents, recording conditions or specific tasks.
The context window grows from 8,000 to 262,144 tokens, a change that permits much longer inputs to be processed in one context. The project also reports improvements to instruction-following, tool use and reasoning. Its training pipeline included unpaired multimodal pretraining, alignment using paired data, and further fine-tuning with reinforcement learning and self-distillation. The available source does not provide detailed benchmark tables or independent evaluations to quantify the reported gains.
Longer Context, Broader Inputs
The release combines two changes that can expand what developers test with open-weight models: processing images and audio and accepting substantially longer inputs. A larger context window may be useful for workflows involving lengthy documents or multiple pieces of source material, while image input could support tasks such as interpreting charts or page layouts. These are potential uses, not guarantees of accuracy in production.
Availability under an Apache 2.0 license matters to teams seeking to inspect, adapt or deploy model weights under the license’s terms rather than rely solely on a hosted proprietary service. The source also says training data and methods are documented. Greater access can support research and customization, but openness by itself does not establish that outputs are reliable, safe or suitable for a particular application.
The project reports stronger results on multilingual translation, legal tasks and reasoning tests, including AIME. Those claims suggest an effort to compete with current open-weight systems across more than text generation. Without scores, test conditions or third-party comparisons in the supplied material, however, readers cannot determine the size or consistency of the gains.
As an affiliate, we earn on qualifying purchases.
From Apertus 1.0 to 1.5
The Apertus project describes its earlier release, Apertus 1.0, as focused on language understanding, with no multimodal input and an 8,000-token context window. Version 1.5 extends that base with image and audio processing, a larger context and additional training intended to improve instruction-following and reasoning.
The new models were trained in several stages, including multimodal pretraining, paired-data alignment and later fine-tuning. The source identifies FineWeb2-HQ+ and Nemotron-CC among the open data sources associated with the project and says it provides information about training data and methodology. The release thus presents transparency and reproducibility as part of the project, alongside model capability.
“The models can interpret documents, diagrams, photos, and spoken language, though speech remains experimental.”
— Apertus project, as described in the supplied release material
audio and image processing software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Performance Still Needs Validation
The available material does not include benchmark scores, evaluation dates, comparisons with named competing models or enough methodology to independently assess the reported performance gains. The statement that Apertus 1.5 outperforms prior models on translation, legal tasks and reasoning tests is a project-reported result, not an independently verified finding in the supplied information.
It is also unclear how the models perform in everyday use, particularly on speech input, complex visual material and very long prompts. The source identifies speech as experimental and gives no detailed results for robustness, accuracy across languages or resistance to unsafe outputs. Training transparency and alignment steps do not settle questions about misuse, bias or suitability for high-stakes decisions.
The source states that the models are available under Apache 2.0, but it does not specify hardware requirements, inference speed, deployment costs or any restrictions tied to particular components. Those details may affect whether researchers and organizations can use the models in practice.
As an affiliate, we earn on qualifying purchases.
Documentation and Further Testing
The project says it plans to continue training and refine safety measures, conduct more evaluation in real-world settings, and gather feedback from the community. It also points to expanded documentation and tools as planned support for researchers and developers. No timetable for those releases is provided in the source material.
For now, the next useful evidence will be detailed model documentation and independently reproducible evaluations that spell out benchmark versions, prompts, scoring methods and comparison baselines. Further testing of speech and other multimodal tasks should help clarify where Apertus 1.5 is dependable and where its experimental features need improvement.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is new in Apertus 1.5?
The release adds image and audio input, increases the context window from 8,000 to 262,144 tokens, and reports improvements in reasoning, instruction-following and tool use.
Which Apertus 1.5 model sizes are available?
The announcement covers 8B and 70B models. The supplied material says they were released on Hugging Face under an Apache 2.0 license.
Can Apertus 1.5 understand speech?
The models include audio input and are described as capable of processing spoken language, but the project characterizes speech understanding as experimental. The source gives no detailed accuracy results.
Are the claimed performance gains independently verified?
Not in the supplied material. The project reports gains on translation, legal and reasoning evaluations, including AIME, but the source does not provide full scores or independent comparisons.
What should users watch for next?
Watch for the project’s planned documentation, tools and real-world evaluations, as well as more detail about safety measures and the performance of multimodal features.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Evergreen bestsellers Picks
bestsellers
As an affiliate, we earn on qualifying purchases.
