🔍 Read the full analysis: What Changed In Apertus 1.5? 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, announced 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 models are available under an Apache 2.0 license, but speech understanding remains experimental and independent real-world performance is not established in the supplied material.
The Apertus project announced version 1.5 on July 24, 2026, adding image and audio input, a 262,144-token context window and reported improvements in reasoning, instruction-following and tool use to its 8B and 70B models, as detailed in the original release notes. The models are available on Hugging Face under an Apache 2.0 license, giving developers access to their weights and, according to the release account, information about training data and methods.
Apertus’s architectural template expands on the previous models through additional training, including 4 trillion multimodal tokens for the 8B model and 2 trillion for the 70B model. The source account gives total training volumes of 19 trillion tokens for the 8B model and 17 trillion for the 70B model. It describes a staged process involving unpaired multimodal pretraining, alignment on paired data, and later fine-tuning that included reinforcement learning and self-distillation.
The models process images and audio alongside text by representing these inputs as tokens in the same sequence. The account identifies Emu3.5 for image tokenization and WavTokenizer for audio. This is intended to support tasks involving documents, diagrams, photographs and spoken language; however, the source specifically says speech understanding remains experimental, so its readiness for routine use should not be assumed.
The context window rises from 8,000 to 262,144 tokens, allowing much longer inputs to be supplied in one interaction. The release account also reports improvements in instruction-following and tool use, and says evaluation results were stronger than those of earlier Apertus models on multilingual translation, legal tasks and reasoning tests including AIME. It does not provide scores or detailed test conditions in the supplied material, so those performance claims cannot be independently compared here.
A Wider Range of Open-Model Tasks
The release brings image and audio input and a much longer context window to models whose weights are available under a permissive license. For developers and researchers, that combination may make it easier to test document analysis, long-input workflows and multimodal applications without relying solely on hosted proprietary systems. The license and disclosed training information also offer a basis for inspecting, adapting and running the models, subject to the project’s documentation and the practical requirements of deployment.
Those capabilities do not by themselves establish that Apertus 1.5 matches leading commercial systems in everyday use. The benchmark claims in the source are not accompanied by scores, independent evaluations or deployment comparisons. Its importance for readers is therefore both practical and conditional: the available capabilities have broadened, while their quality, reliability and cost in real applications remain to be assessed.
Open access also means safety cannot be judged from a license or a description of alignment work alone. The project says it used evaluation and training steps aimed at safety and adherence to the Apertus Charter, but the supplied account does not detail independent audits or observed misuse. Developers considering deployment will need to evaluate behavior for their own users and tasks.
multimodal AI models with image and audio input
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From Text Models to Multimodal Input
The earlier Apertus release is described as focused on language, with an 8,000-token context limit and no comparable image or audio input. Version 1.5 changes those boundaries through additional multimodal training and a revised approach to representing different input types. The update is not presented as a new model size: it continues to use 8B and 70B architectures.
The project’s stated training sources include FineWeb2-HQ+ and Nemotron-CC, alongside multimodal data. According to the source account, the team emphasizes transparency about data and methods and released the models’ weights on Hugging Face. It also describes alignment to the Apertus Charter and a training pipeline combining pretraining, paired-data alignment and fine-tuning. Those details help explain the release’s design goals, but do not substitute for full technical documentation or external evaluation.
The announcement also reflects a broader move among open-weight model projects toward handling more than text and accepting longer prompts. In Apertus 1.5, the notable change is the combination of multimodal input and extended context in the project’s existing two model sizes. Whether that combination translates into dependable results depends on performance testing beyond the release claims.
“The models are released to Hugging Face under an Apache 2.0 license.”
— Apertus project, as summarized by ThorstenMeyerAI.com
Performance and Speech Still Need Testing
The supplied source does not include benchmark tables, evaluation procedures, model cards or independent test results. It is therefore unclear how large the reported gains are, how the models compare with current alternatives under matched conditions, or how performance varies across languages and task types. The claim that the models move closer to frontier open-weight systems is an assessment in the source, not a result that can be verified from the figures provided.
Audio performance is explicitly experimental, and the material does not specify supported audio formats, languages, accuracy or limitations. Nor does it give detailed measurements of safety, robustness, latency, hardware needs or inference costs. Those factors will affect whether the models suit practical deployments, particularly for sensitive documents or high-stakes tasks.
The source says the team is working on safety and alignment, but does not identify independent audits or provide results from real-world trials. The extent to which released training data and methods are documented is also not fully detailed here. These questions remain open as researchers and developers test the models.
Documentation and Community Testing
The source says the Apertus team plans to continue training and evaluation, with further work on safety measures and domain-specific fine-tuning. It also points to expanded real-world evaluation and community feedback as part of the next phase. More complete documentation and tools are described as planned, but no firm publication dates are given in the supplied account.
For users, the next useful evidence will be reproducible benchmark results, clearer technical guidance and testing of image and audio tasks outside curated evaluations. The project also plans to explore improvements to speech recognition and reasoning. Until those updates and broader evaluations are available, speech should be treated as experimental and performance claims as release-reported rather than independently established.
Key Questions
What are the main changes in Apertus 1.5?
Apertus 1.5 adds image and audio input, expands the context window from 8,000 to 262,144 tokens, and reports improvements in reasoning, instruction-following and tool use. It retains the 8B and 70B model sizes.
Can Apertus 1.5 understand speech?
The source says the models can process audio, but describes speech understanding as experimental. It does not give accuracy results, supported-language details or a complete account of audio limitations.
Where are the models available, and what license do they use?
The models were released on Hugging Face under the Apache 2.0 license, according to the source account. Users should check the project’s release materials for the applicable terms, files and instructions.
How much better are the models than earlier Apertus versions?
The source reports gains on multilingual translation, legal tasks and reasoning tests such as AIME, but provides no scores or testing details. The size and consistency of those improvements cannot be established from the supplied information.
What should developers watch for next?
The project says it plans more safety work, domain-specific fine-tuning, expanded evaluations and additional documentation and tools. No timetable is provided for those steps in the source material.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
