---

license: gemma

base_model: google/gemma-3-4b-it

pipeline_tag: image-text-to-text

tags:

- arxiv:2110.14168

- arxiv:2311.12022

- gemma3

- vision

- arxiv:2203.10244

- arxiv:2210.03057

- gguf

- arxiv:1903.00161

- arxiv:1907.10641

- arxiv:1705.03551

- safetensors

- arxiv:2009.03300

- text-generation-inference

- arxiv:1911.01547

- arxiv:1905.10044

- arxiv:2106.03193

- arxiv:2502.12404

- arxiv:2104.12756

- arxiv:2502.21228

- vlm

- transformers

- arxiv:1911.11641

- arxiv:1810.12440

- arxiv:2103.03874

- arxiv:1908.02660

- arxiv:2107.03374

- multimodal

- image-text-to-text

- eval-results

- arxiv:1904.09728

- conversational

- base_model:google/gemma-3-4b-pt

- arxiv:2404.12390

- arxiv:2312.11805

- arxiv:1910.11856

- arxiv:1905.07830

- base_model:finetune:google/gemma-3-4b-pt

- arxiv:2311.16502

- arxiv:2304.06364

- region:us

- license:gemma

- arxiv:2108.07732

- quantized

- arxiv:2404.16816

language:

- en

---

<div align="center">

gemma-3-4b-it — GGUF Quantizations (VLM)




Quantized GGUF versions of google/gemma-3-4b-it

This is a Vision-Language Model (VLM) — it can understand both text and images.

Works with llama.cpp · LM Studio · Jan · Ollama

Quantized by Dhptl on June 11, 2026 using quant-kit

</div>

---

> [!IMPORTANT]

> This VLM requires TWO files — a text backbone GGUF and the mmproj vision encoder GGUF.

> Download one text backbone (e.g. Q4_K_M) and the mmproj file. Both must be in the same folder.

---

📦 Available Files

🔤 Text Backbone (quantized — pick ONE)

| Filename | Size | RAM Required | Quant | Quality | Best For |

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

| gemma-3-4b-it-Q2_K.gguf | 1.61 GB | ~3.1 GB | Q2_K | ⭐ | Extreme compression, significant quality loss. |

| gemma-3-4b-it-Q3_K_L.gguf | 2.08 GB | ~3.6 GB | Q3_K_L | ⭐⭐⭐ | Slightly better than Q3_K_M, still a compromise. |

| gemma-3-4b-it-Q3_K_M.gguf | 1.95 GB | ~3.5 GB | Q3_K_M | ⭐⭐⭐ | Very small file. Quality drop noticeable. |

| gemma-3-4b-it-Q3_K_S.gguf | 1.80 GB | ~3.3 GB | Q3_K_S | ⭐⭐ | Very high compression, high quality loss. |

| gemma-3-4b-it-Q4_K_M.gguf | 2.32 GB | ~3.8 GB | Q4_K_M ✅ Recommended | ⭐⭐⭐⭐ | Best balance of size and quality. Recommended for most users. |

| gemma-3-4b-it-Q4_K_S.gguf | 2.21 GB | ~3.7 GB | Q4_K_S | ⭐⭐⭐½ | Good speed/size balance, slight quality loss. |

| gemma-3-4b-it-Q5_K_M.gguf | 2.64 GB | ~4.1 GB | Q5_K_M | ⭐⭐⭐⭐½ | Better quality than Q4, slightly larger. Great if you have the RAM. |

| gemma-3-4b-it-Q5_K_S.gguf | 2.57 GB | ~4.1 GB | Q5_K_S | ⭐⭐⭐⭐ | Large but accurate. |

| gemma-3-4b-it-Q6_K.gguf | 2.97 GB | ~4.5 GB | Q6_K | ⭐⭐⭐⭐⭐ | Near-perfect quality, very large. |

| gemma-3-4b-it-Q8_0.gguf | 3.85 GB | ~5.3 GB | Q8_0 | ⭐⭐⭐⭐⭐ | Closest to original quality. Use when RAM is not a concern. |

🖼️ Vision Encoder — mmproj (always required, always F16)

| Filename | Size | Notes |

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

| gemma-3-4b-it-mmproj-f16.gguf | 0.79 GB | Always F16 — vision encoder is not quantized |

> ⚠️ You need BOTH files — one text backbone + the mmproj — to run this VLM.

---

⚡ Speed Benchmarks

Run python benchmark.py --model gemma-3-4b-it to generate results.

---

🚀 How to Use

LM Studio (Easiest — GUI)

1. Search for Dhptl/gemma-3-4b-it in LM Studio
2. Download the Q4_K_M text file and the mmproj file
3. Load the model — LM Studio automatically uses both files

Ollama

ollama run dhptl/gemma-3-4b-it

llama.cpp CLI — Text + Image

# Download both files to the same directory, then:
./llama-llava-cli \
  -m gemma-3-4b-it-Q4_K_M.gguf \
  --mmproj gemma-3-4b-it-mmproj-f16.gguf \
  --image /path/to/your/image.jpg \
  -p "Describe this image in detail." \
  -n 512

llama.cpp CLI — Text only (no image)

./llama-cli \
  -m gemma-3-4b-it-Q4_K_M.gguf \
  -p "You are a helpful assistant." \
  --conversation

Python — llama-cpp-python

from llama_cpp import Llama
from llama_cpp.llama_chat_format import Llava16ChatHandler

# Load VLM with mmproj
chat_handler = Llava16ChatHandler(clip_model_path="./gemma-3-4b-it-mmproj-f16.gguf")
llm = Llama(
    model_path="./gemma-3-4b-it-Q4_K_M.gguf",
    chat_handler=chat_handler,
    n_gpu_layers=-1,
    n_ctx=4096,
    logits_all=True,
)

# Text + image inference
response = llm.create_chat_completion(
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "image_url", "image_url": {"url": "https://example.com/image.jpg"}},
                {"type": "text",      "text":      "What do you see in this image?"}
            ]
        }
    ]
)
print(response["choices"][0]["message"]["content"])

---

🔍 VLM Architecture

This model uses a two-component architecture:

| Component | File | Purpose |

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

| Text Backbone | gemma-3-4b-it-Q4_K_M.gguf | Language understanding & generation |

| Vision Encoder (mmproj) | gemma-3-4b-it-mmproj-f16.gguf | Image feature extraction (always F16) |

> Why is mmproj always F16?

> The vision encoder maps image pixels to token embeddings. Quantizing it causes

> visible visual artifacts and degraded image understanding. It stays at F16 (half precision)

> which is already very efficient at ~1-2GB for most models.

---

🔍 About GGUF Quantization

| Format | Bits/weight | Quality |

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

| Q3_K_M | ~3.3 | ⭐⭐⭐ |

| Q4_K_M | ~4.5 | ⭐⭐⭐⭐ ← recommended |

| Q5_K_M | ~5.6 | ⭐⭐⭐⭐½ |

| Q8_0 | ~8.5 | ⭐⭐⭐⭐⭐ |

---

💬 Community & Feedback

Found an issue? Open a Discussion in the Community tab.

If useful, please:

• ⭐ Star quant-kit on GitHub
• 👍 Like this model on HuggingFace