---

license: apache-2.0

language:

• en
• zh

base_model:

• BytedanceDouyinContent/SAIL-VL-2B

pipeline_tag: image-text-to-text

tags:

• gguf
• vision-language
• multimodal
• image-text-to-text
• llama.cpp
• qwen2
• internvit
• mmproj
• q4_k_m
• q8_0
• 2b
• ocr
• chart-qa

---

SAIL-VL-2B GGUF






Hugging Face release: ShayonSarker/SAIL-VL-2B-GGUF

Kaggle notebook: Run the complete conversion notebook

GitHub source: Dadhichi-Sarker-Shayon/SAIL-VL-2B-GGUF

Unofficial GGUF conversion of BytedanceDouyinContent/SAIL-VL-2B, packaged for inference with llama.cpp.

The repository contains the Qwen2 text model in F16, Q8_0, and Q4_K_M formats, plus the InternViT multimodal projector required for image prompts.

This GitHub repository contains the reproducible Kaggle conversion pipeline and release metadata. The multi-gigabyte GGUF artifacts are hosted on Hugging Face at ShayonSarker/SAIL-VL-2B-GGUF.

Repository contents

| File | Purpose |

|---|---|

| SAIL-VL-2B-GGUF.ipynb | End-to-end CPU/Kaggle conversion, validation, and upload notebook |

| REPRO.json | Conversion commit, inference output, sizes, and short hashes |

| SHA256.txt | Full checksums for the published GGUF artifacts |

| LICENSE | Apache License 2.0 |

Published artifacts

Download the binaries from the Hugging Face release:

• sail-vl-2b-f16.gguf
• sail-vl-2b-Q8_0.gguf
• sail-vl-2b-Q4_K_M.gguf
• mmproj-sail-vl-2b-f16.gguf

Format status

| File | Status | Purpose |

|---|---|---|

| sail-vl-2b-f16.gguf | Published | Reference text model |

| sail-vl-2b-Q8_0.gguf | Published | Balanced text quantization |

| sail-vl-2b-Q4_K_M.gguf | Published | Smaller text quantization |

| mmproj-sail-vl-2b-f16.gguf | Published | Vision projector for image input |

| mmproj Q8_0 / Q4_K_M | Not published | Only the F16 projector was converted and validated |

The vision projector is only 628 MB in F16, so quantising it further is not worth the validation risk. Any text GGUF from this repository can be paired with the F16 mmproj.

Files

| File | Approx. size | Purpose |

|---|---:|---|

| sail-vl-2b-f16.gguf | 3560.4 MB | Reference text model |

| sail-vl-2b-Q8_0.gguf | 1894.5 MB | Balanced text quantization |

| sail-vl-2b-Q4_K_M.gguf | 1117.3 MB | Smaller text quantization |

| mmproj-sail-vl-2b-f16.gguf | 628.2 MB | Vision projector for image input |

Use the mmproj file with any text GGUF from this repository. The validation run below used sail-vl-2b-Q8_0.gguf.

Rebuild the conversion

Run the published Kaggle notebook, or open SAIL-VL-2B-GGUF.ipynb locally. Enable Internet, select the CPU accelerator, add HF_TOKEN as a Kaggle secret, and run all cells. The notebook downloads the source model, builds llama.cpp, converts the text and vision projector, runs the semantic gate, writes checksums, and uploads the release to Hugging Face.

Generated GGUFs and local build directories are intentionally ignored by Git; download them from the Hugging Face release instead.

Quick start

Download the files with the Hugging Face Hub:

python -c "from huggingface_hub import snapshot_download; snapshot_download('ShayonSarker/SAIL-VL-2B-GGUF', local_dir='SAIL-VL-2B-GGUF')"

Build a current llama.cpp checkout with llama-mtmd-cli:

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release --target llama-mtmd-cli llama-cli

Run an image prompt:

./llama.cpp/build/bin/llama-mtmd-cli \
  -m ./SAIL-VL-2B-GGUF/sail-vl-2b-Q4_K_M.gguf \
  --mmproj ./SAIL-VL-2B-GGUF/mmproj-sail-vl-2b-f16.gguf \
  --image ./image.jpg \
  -p "Describe the image and read any visible text." \
  --temp 0 -n 256

For text-only prompts, use llama-cli with one of the text GGUFs; the mmproj file is not needed.

Validation

The conversion was tested with llama.cpp commit fc343a84bbd925b37dde3219de35ea0bed50d630 on CPU. All GGUF files passed the GGUF magic check, and the Q8_0 model passed a semantic image test using the source repository's statics/14.jpg. Verbatim model output:

> The image features a French Bulldog lying on a worn, turquoise wooden deck. ... The text overlaid on the image reads "MONDAY. JUST...MONDAY."

This is the only vision release in this collection with a passing image gate; the SmolVLM-500M multimodal attempt scored 0/50 and was not published.

The exact inference output, file sizes, conversion commit, and short hashes are recorded in REPRO.json. Full SHA-256 checksums are in SHA256.txt.

The current build may print a non-fatal tokenizer metadata warning about token 128247 (</s>). It did not prevent inference or the semantic validation gate.

License and attribution

The source model is distributed under the Apache License 2.0. See LICENSE and the original model card for upstream details, training information, and citation.

This repository is a community conversion and is not an official ByteDance release.