🔍 Read the full analysis: Apertus 1.5 Release Notes: What’s New 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, released July 24, 2026, adds image and audio input to the project’s 8B and 70B models and expands their context window from 8,000 to 262,144 tokens. The release notes describe improved reasoning, instruction-following and tool use, but provide no benchmark figures here; speech remains experimental.
The Apertus project released Apertus 1.5 on July 24, 2026, updating its 8B and 70B models with image and audio input and expanding the context window from 8,000 to 262,144 tokens. The changes broaden what the open-weight models can process, although the release material describes speech understanding as experimental.
The project says Apertus 1.5 was built through additional pretraining, adding 4 trillion multimodal tokens for the 8B model and 2 trillion for the 70B model. The stated cumulative training totals are 19 trillion tokens for 8B and 17 trillion for 70B. The source material does not provide a detailed breakdown of the added data or a direct account of how performance changes with each model size.
Images and audio are represented as tokens alongside text. The release notes name the Emu3.5 tokenizer for images and WavTokenizer for audio. The models are described as able to interpret material such as documents, diagrams and photos, as well as spoken language; speech functionality is still characterized as experimental.
Apertus 1.5 is available through Hugging Face under the Apache 2.0 license, according to the source. The project describes the release as including open weights and information about training data and methodology. Its training process included multimodal pretraining, alignment using paired data, and fine-tuning with reinforcement learning and self-distillation. The release notes also report gains on multilingual translation, legal tasks and reasoning tests, including AIME, but do not state scores, testing conditions or comparisons in the supplied material.
More Input, Longer Model Context
The expanded context window lets a model take in substantially more material in a single interaction than the previous 8,000-token limit. That may be useful for tasks involving long documents or several connected pieces of information, while image input could help with documents, diagrams and photographs. These are potential uses, not proof of performance in particular workflows.
For researchers and developers, the release combines multimodal input and open weights under a license that permits broad use. That can support experimentation and adaptation outside a closed service. The practical value will depend on factors not resolved by the release summary, including real-world accuracy, hardware requirements, safety behavior and the quality of its handling of different languages and image types.
The reported benchmark improvements could make the models more useful for reasoning-heavy work, but the supplied source gives no numerical results or independent evaluation details. Readers should treat claims of improved performance as project-reported results until comparable evaluation data are available.
As an affiliate, we earn on qualifying purchases.
From Apertus 1.0 to Multimodal Input
The source describes Apertus 1.0 as a language-focused release with an 8,000-token context window and no multimodal capabilities. Version 1.5 builds on those models through further training and changes to how image and audio information is processed. The update is therefore a continuation of the same model family, not a separate model series.
The project presents openness and transparency as central features, citing public data sources including FineWeb2-HQ+ and Nemotron-CC, as well as documentation about training methods. The material supplied for this report does not include links to the full technical documentation or underlying benchmark tables, so those details cannot be independently assessed here.
“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 input development board
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Speech and Benchmark Evidence
Speech understanding remains experimental, and the supplied release summary does not specify its limitations, supported languages or performance in everyday use. It also does not give benchmark scores, baselines, test conditions or independent evaluations for the reported improvements in reasoning, translation and legal tasks.
The material does not provide detailed figures for inference speed, memory use or hardware needs, nor does it establish how reliably the models handle varied images and long documents. The project cites alignment and evaluation work, but real-world safety and robustness remain questions for further testing. Open weights also mean that downstream use will depend on how users deploy and modify the models.
As an affiliate, we earn on qualifying purchases.
Further Testing and Project Updates
The project says it plans to continue refining Apertus with additional training, safety measures and domain-specific fine-tuning. It also says it intends to broaden evaluations, particularly in real-world settings, and use community feedback to guide development. These are stated plans; the supplied material does not give dates for the next release or documentation milestones.
For users assessing whether to adopt the models, the next useful evidence will include fuller technical documentation, benchmark results with clear comparison baselines, and independent tests of speech, safety and practical deployment. Until then, Apertus 1.5’s described capabilities should be distinguished from validated results in specific applications.
As an affiliate, we earn on qualifying purchases.
Key Questions
When was Apertus 1.5 released?
The source material says Apertus 1.5 was released on July 24, 2026.
What changed in Apertus 1.5?
The update 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. Speech understanding is still described as experimental.
Which model sizes are available?
The release covers 8B and 70B models. The supplied material does not give further model specifications or deployment requirements.
Where can developers access the models?
The source says the models are available on Hugging Face under the Apache 2.0 license, with open weights and information about training data and methodology.
Are the reported performance gains independently verified?
The supplied release summary reports gains on several benchmark areas but provides no scores, testing conditions or independent results. The strength of those comparisons cannot be established from the material provided.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
