Install from WinGet (Windows)
winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Hansda/ScreenVLM-GGUF:F16# Run inference directly in the terminal:
llama cli -hf Hansda/ScreenVLM-GGUF:F16Build from source code
git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf Hansda/ScreenVLM-GGUF:F16# Run inference directly in the terminal:
./build/bin/llama cli -hf Hansda/ScreenVLM-GGUF:F16Use Docker
docker model run hf.co/Hansda/ScreenVLM-GGUF:F16ScreenVLM GGUF
F16 and Q8_0 conversions of the ScreenParse-v2 checkpoint, with a shared F16 vision encoder/projector. Q8_0 is derived from the original-weight F16 conversion.
| File | Component | Quantization | Size | Status |
|---|---|---|---|---|
| ScreenVLM-Q8_0.gguf | decoder | Q8_0 | 239,333,536 bytes | Available |
| ScreenVLM-F16.gguf | decoder | F16 | 447,246,496 bytes | Available |
| mmproj-ScreenVLM-F16.gguf | vision | F16 | 190,031,392 bytes | Available |
Usage
Use llama.cpp at 0ec4027:
llama-server -m ScreenVLM-Q8_0.gguf --mmproj mmproj-ScreenVLM-F16.gguf \
-ngl 99 -c 4096 -np 1 --jinja --special \
--host 127.0.0.1 --port 8080
Send one screenshot followed by Generate the screen representation for this UI:.
--special retains the <loc_ control prefix. ScreenTag bounding boxes
use integer coordinates in [0,500]. For CPU inference, use
-ngl 0 --no-mmproj-offload.
Preserve the native Idefics3 processing: longest edge 2048, 512-pixel tiles plus overview, 4× pixel shuffle, Lanczos, and mean/std 0.5. The original chat template is embedded and supplied separately. The converter copy removes two token IDs outside the source weight tables; all model-representable token IDs and source weights are retained. See tokenizer-repair.json and CHANGES.md.
Source and license
Apache-2.0, as declared by the source and its named components. Complete terms and source attribution are included in LICENSE and ATTRIBUTION.md. Commands, pinned provenance and hashes: REPRODUCE.md, runtime-info.json, model-manifest.json.
- Downloads last month
- -
8-bit
16-bit
Model tree for Hansda/ScreenVLM-GGUF
Base model
docling-project/ScreenVLM
Install (macOS, Linux)
# Start a local OpenAI-compatible server with a web UI: llama serve -hf Hansda/ScreenVLM-GGUF:F16# Run inference directly in the terminal: llama cli -hf Hansda/ScreenVLM-GGUF:F16