---

base_model: HuggingFaceTB/SmolVLM-500M-Instruct

tags:

• gguf
• llama.cpp
• vision
• smolvlm
• chartqa

license: apache-2.0

pipeline_tag: image-text-to-text

---

SmolVLM-500M-Instruct full-multimodal GGUF (Q4 text + F16 mmproj)

Base: HuggingFaceTB/SmolVLM-500M-Instruct (Apache-2.0) · Text tower: Q4_K_M imatrix quant (reused from our measured build, +1.85% text PPL) · Vision: F16 mmproj via convert_hf_to_gguf.py --mmproj

Status

Pipeline running. Vision eval (50 real ChartQA images via llama-mtmd-cli) fills the table below on gate pass (bar ≥20%).

Eval (pending gate)

| model | ChartQA-50 vision relaxed-acc |

|---|---|

| SmolVLM-500M-Instruct base (published) | 63.2 (test, for reference) |

| Q4 + mmproj (this repo) | _pending_ |

Use

./llama-mtmd-cli -m q4.gguf --mmproj mmproj-f16.gguf --image chart.png -p "What value is shown for 2015?"

Needs both files from this repo. Fully offline.