Velvet-25B
October 14, 2025
A smart model optimized for text processing
The essence of Velvet-25B
Velvet-25B is a multilingual model, tailored for long text and document processing, trained in a natively multilingual mode, without using English as the reference language.
Languages
Architecture
Velvet-25B is built on a long-context architecture that preserves the model’s attention capabilities, allowing it to process very large texts while maintaining coherence and accuracy across distant sections.
Training Dataset
The training process started with over 15 trillion tokens and ended with more than 7 trillion tokens.
Context Window
128K Tokens.
Thanks to its context window, the model can handle long and complex documents such as legal texts, scientific reports, or legislative acts.
Parameters
25 Billion parameters.
Vocabulary
148K Tokens.
Specialization
6M Instructions.
Data Freshness
Data cutoff date: June 2025.
Capabilities
Instruction Following
Information Extraction
Multistep Dynamic Reasoning
Structured Output Generation
Function
Calling
Machine Translation
Textual
Entailment
Text
Classification
Question Answering
Multiturn Conversation
Text
Completion
Summarization
Paraphrasing
RAG
Main Features
Bespoke Ready
The model is designed for customization and can be specialized for various domains.
Agentic Architecture
Through integration with Agentic and RAG architectures, the model connects in real time to external knowledge sources, keeping contents always up to date.
Agentic Orchestration
Post-training methods like reinforcement learning, refine reasoning and enable agentic orchestration, efficiently selecting the right AI agents for different tasks.
Deployment
Velvet runs efficiently on a single GPU, making AI accessible to both large organizations and those with limited infrastructure.
Optimization
Several optimization techniques let the model adjust processing depth based on request complexity, fully leveraging infrastructure and reaching over 90% GPU efficiency, while cutting energy use and response times.
Edge and Cloud
Velvet models are compatible with edge devices and ready for emerging European GPUs, supporting deployment across large data centers, small infrastructures, and mobile systems.
Why Velvet-25B?
Performance and Scale
Velvet-25B deliver state-of-the-art results relative to their size and operating cost.
Efficiency and Applicability
Velvet models are focused on improving the efficiency and practical use of language models.
Enterprise Adaptation
Velvet 25B is natively integrated into AIWave, a multi-model and multiagentic platform that enables the natural application of Velvet models in a wide range of complex enterprise use cases.