adoslabs/liara-24b

VISIONConcurrent Unit Cost:2Model Size:24BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 23, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Liara 24B by adoslabs is an Italian-first multimodal assistant, built upon Mistral-Small-3.2-24B-Instruct-2506. This 24 billion parameter model is specialized for Italian language with a warm, concise, first-person persona, and supports both text and image modalities. It is designed for general-purpose Italian assistance, including conversation, tool-use scaffolding, and image understanding.

Loading preview...

Liara 24B: Italian-First Multimodal Assistant

Liara 24B is an Italian-first multimodal assistant developed by adoslabs, based on the mistralai/Mistral-Small-3.2-24B-Instruct-2506 model. This 24 billion parameter model has been specialized through a LoRA (knowledge distillation of Liara's persona, tool-use behavior, and safety policy) which was then merged directly into the weights, creating a standalone checkpoint. The vision encoder remains unchanged from the base Mistral model.

Key Capabilities & Features

  • Italian Language Optimization: Specifically designed for Italian, featuring a warm, concise, and first-person persona.
  • Multimodal: Supports both text and image inputs, utilizing the Pixtral vision capabilities from its base model.
  • Mistral Consolidated Format: Shipped in the Mistral consolidated format (consolidated.safetensors), which is crucial for ensuring vision functionality when serving with tools like vLLM.
  • General-Purpose Assistant: Intended for conversational AI, tool-use scaffolding, and image understanding in Italian contexts.

Usage Considerations

  • Serving: Requires specific vllm configurations (--tokenizer-mode mistral, --config-format mistral, --load-format mistral) to enable vision capabilities.
  • Image Input: For documents like PDFs or Excel files, pages must be rendered to images before being sent to the model.
  • Limitations: Trained largely on synthetic/curated Italian data. Like other LLMs, it can hallucinate and its outputs should be validated. Its persona and policies are specific to the Liara product and may not generalize to other framings.

License

Liara 24B is a derivative of Mistral-Small-3.2-24B-Instruct-2506 and is released under the Apache-2.0 license.

Finetunes
1 models