AMALIA-VL-DPO-GGUF

AMALIA-VL-DPO — GGUF Quantizations

GGUF quantizations of amalia-llm/AMALIA-VL-DPO, the open-source vision-language model targeting European Portuguese (pt-PT), developed by a consortium of Portuguese universities and research centres and funded by the Government of Portugal.

The original model uses a LLaVA-NeXT architecture with a SigLIP vision encoder (1152 hidden dim, 384px image size, patch size 16) and a LLaMA-based language backbone (9.15B parameters, 42 layers, 32768 context). The base model was trained in 3 stages: modality alignment, visual instruction following, and direct preference optimization (DPO).

Note: These are community-converted GGUF files. The original model was released by the AMALIA LLM Project under the Apache 2.0 license. All quantizations were produced from the official BF16 safetensors checkpoint converted to F16 GGUF, then quantized using llama.cpp.

---

Available Files

The mmproj-model-f16.gguf file is shared across all quantizations — you only need one copy regardless of which quant you run.

---

Usage

With llama-mtmd-cli (recommended for local inference)

llama-mtmd-cli \
  -m AMALIA-VL-DPO-Q4_K_M.gguf \
  --mmproj mmproj-model-f16.gguf \
  -ngl 99 \
  -fa on \
  --image /path/to/image.jpg \
  -p "Descreve esta imagem em português"

Interactive chat mode (no image or prompt required at launch):

llama-mtmd-cli \
  -m AMALIA-VL-DPO-Q4_K_M.gguf \
  --mmproj mmproj-model-f16.gguf \
  -ngl 99 \
  -fa on

With video:

llama-mtmd-cli \
  -m AMALIA-VL-DPO-Q4_K_M.gguf \
  --mmproj mmproj-model-f16.gguf \
  -ngl 99 \
  --video /path/to/video.mp4 \
  -p "O que acontece neste vídeo?"

With llama-server (OpenAI-compatible API)

llama-server \
  -m AMALIA-VL-DPO-Q4_K_M.gguf \
  --mmproj mmproj-model-f16.gguf \
  -ngl 99 \
  -fa on \
  --host 0.0.0.0 \
  --port 8080

Recommended sampling parameters

The original model uses ChatML format with the following recommended settings:

--chat-template chatml \
--temp 0.7 \
--top-p 0.9 \
--repeat-penalty 1.1

---

VRAM Requirements

Context length also consumes VRAM. The above estimates are for the model weights + mmproj at short context. KV cache adds approximately 1–5 GB depending on context length used.

---

Conversion Details

The GGUFs were produced from the official amalia-llm/AMALIA-VL-DPO checkpoint using llama.cpp.

Process:

1. The LLM backbone was extracted from the LLaVA-NeXT wrapper using transformers and converted to F16 GGUF via convert_hf_to_gguf.py
2. The SigLIP vision encoder and MLP projector were extracted using llava_surgery_v2.py and converted to an mmproj GGUF via convert_image_encoder_to_gguf.py --clip-model-is-siglip
3. All quantizations were produced from the F16 GGUF master using llama-quantize

Vision encoder specs:

• Architecture: SigLIP (siglip_vision_model)
• Hidden size: 1152
• Image size: 384×384
• Patch size: 16
• Sequence length: 576 tokens per image

---

About AMALIA-VL

AMALIA-VL is developed by a consortium of Portuguese universities and research centres including NOVA University Lisbon, Instituto Superior Técnico, the University of Coimbra, the University of Porto, the University of Minho, and the Foundation for Science and Technology (FCT), funded by the Government of Portugal.

Training was carried out on the MareNostrum5 supercomputer and the DEUCALION supercomputer using 64 NVIDIA H100 GPUs across three training stages.

For full details, refer to the technical report and the official model card.

---

Citation

If you use these GGUF files in your work, please cite the original AMALIA-VL paper:

@article{gloria2026amalia,
    title={AMALIA-VL: A Native European Portuguese Open-Source Vision and Language Model},
    author={Glória-Silva, Diogo and Cardeira, João and da Luz, Manuel Letras and Simplício, Afonso and Vinagre, Gonçalo and Tavares, Diogo and Ferreira, Rafael and Calvo, Inês and Vieira, Inês and Semedo, David and others},
    journal={arXiv preprint arXiv:2606.19100},
    year={2026}
}