The base model was specialized with a LoRA (knowledge distillation of Liara's persona, tool-use behaviour and safety policy) and then merged into the weights — this is a standalone merged checkpoint, there is no separate adapter to attach. The vision encoder is unchanged from the base (the LoRA only touched the language model).

Language: optimized for Italian (warm, concise, first-person persona).
Modalities: text and images (Pixtral vision, same as the base).
Format: Mistral consolidated (consolidated.safetensors + params.json + tekken.json).

⚠️ Serving — the Mistral format is required for vision

On vLLM, Mistral-Small-3.2 vision works only with the full Mistral format. Loading the HF-shard layout makes the vision encoder produce garbage (see vLLM #20025). This repo ships the consolidated weights precisely so vision works out of the box:

bash

vllm serve adoslabs/liara-24b \
  --tokenizer-mode mistral \
  --config-format mistral \
  --load-format mistral \
  --served-model-name liara \
  --limit-mm-per-prompt '{"image": 4}'

Then call it with the OpenAI-compatible API:

python

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

# text
client.chat.completions.create(model="liara", messages=[
    {"role": "user", "content": "Ciao Liara, cosa mi consigli di cucinare stasera?"}])

# image + text
client.chat.completions.create(model="liara", messages=[{"role": "user", "content": [
    {"type": "text", "text": "Cosa c'è in questa foto?"},
    {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,..."}}]}])

Documents (PDF/Excel): the model sees pixels, not files — render pages to images before sending them.

Intended use

General-purpose Italian assistant: conversation, tool-use scaffolding, image understanding. Not a safety-critical system; validate outputs before acting on them.

License

Derivative of Mistral-Small-3.2-24B-Instruct-2506, released under Apache-2.0. The merged weights are redistributed under the same license.

Limitations

Trained largely on synthetic / curated Italian data. Like any LLM it can hallucinate; do not treat its output as authoritative. It reflects the persona and policies of the Liara product and may not generalize to other framings.