m-polignano/ANITA-NEXT-24B-Magistral-2506-ITA

TEXT GENERATIONPricing:Input $0.7 / Cached $0.04 / Output $1.16Concurrent Unit Cost:2Model Size:24BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 8, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

ANITA-NEXT-24B-Magistral-2506-ITA by m-polignano is a 24 billion parameter multilingual (English + Italian) thinking model, fine-tuned from Magistral-Small-2506 (a Mistral-based model). It is designed to provide an improved model for Italian language use cases, serving as a general-purpose foundation for further task-specific fine-tuning. The model supports a context length of 32768 tokens, though performance may degrade after 40k tokens, and is optimized for general language understanding and generation in both English and Italian.

Loading preview...

ANITA-NEXT-24B-Magistral-2506-ITA Overview

ANITA-NEXT-24B-Magistral-2506-ITA is a 24 billion parameter "Thinking Model" from the ANITA (Advanced Natural-based interaction for the Italian language) family, developed by Ph.D. Marco Polignano and the SWAP Research Group at the University of Bari Aldo Moro, Italy. This model is a fine-tuned version of Magistral-Small-2506, which itself is based on the Mistral architecture. It is specifically designed as a multilingual model supporting both English and Italian, with a primary focus on enhancing capabilities for Italian language use cases.

Key Capabilities

  • Multilingual Support: Optimized for both English and Italian language processing.
  • General Purpose: Serves as a foundational model for various NLP tasks, intended for further fine-tuning on specific Italian tasks.
  • Fine-tuning Methodology: Utilizes QLoRA 4-bit for supervised fine-tuning (SFT) on instruction-based datasets, and DPO (Direct Preference Optimization) with the mlabonne/orpo-dpo-mix-40k dataset for alignment with human preferences.
  • Context Length: Supports a context window of up to 128k tokens, though performance degradation is noted after 40k tokens.
  • Efficient Deployment: Can run on a single GPU with 19.56GB of VRAM using 4-bit quantization, as demonstrated in a provided Colab notebook.

Good For

  • Italian Language Applications: Ideal for researchers and developers focusing on Italian NLP tasks and applications.
  • Foundation for Customization: Suitable as a base model for further specialized fine-tuning to meet specific project requirements.
  • Research and Development: Released for research purposes, offering a robust platform for exploring advanced natural language interactions in Italian.

Finetunes
2 models