mysnipbox

Models & Setup

Local models (Ollama)

Run vision and text models on your own machine, a remote Ollama server, or your own GPU server, and what hardware you realistically need.

Local models run entirely on your computer through Ollama, so your images never leave the device. They're a good fit for privacy, but be realistic about hardware. Read the requirements below before you start.

Hardware guide (rough)

SetupWhat to expect
32GB+ unified memory, or a GPU with roughly 16GB+ VRAMRecommended: good quality and workable speed with mid-size vision models.
16GB RAM / roughly 8GB VRAM or lessExperimental: small models only; slower and less accurate.
CPU onlyNot recommended: very slow.

What you actually need depends on the model. This is only a rough guide.

1. Install Ollama

  1. 1Download the installer for macOS or Windows from ollama.com/download and run it.
  2. 2Open a terminal and run ollama --version to confirm it works.
  3. 3Ollama runs a local server in the background at http://localhost:11434. That's the endpoint mysnipbox connects to.

2. Choose and download models

mysnipbox uses two kinds of model with different jobs, and you can set them separately:

  • Vision model: looks at the image and describes it. Needs the most memory. Pick one from the verified list below.
  • Text model: handles tag translation and embeddings (tag matching). Much lighter. Pick a translation model from the second list; nomic-embed-text works for embeddings.

Verified vision models

These four vision models are the ones built into the app's model picker.

ModelSizeMin. recommended memoryFeatures & useOllama command
Moondream~2.0 GB8GB RAMUltra-light. Fast tag extraction on low-spec machines.ollama run moondream
Qwen2.5-VL 3B~3.2 GB8GB RAM / VRAMA light vision model tuned for everyday laptops.ollama run qwen2.5vl:3b
MiniCPM-V 2.6 (8B)~5.5 GB16GB+ RAMRecommended: the best balance of speed and analysis quality.ollama run minicpm-v:latest
Llama 3.2 Vision (11B)~7.9 GB+24GB~32GB+ RAMHigh-quality, in-depth analysis for machines with lots of memory.ollama run llama3.2-vision

Tag translation models

To translate tags locally, a small text model is enough. Pick one by the languages you need:

ModelSizeMin. recommended memoryFeatures & useOllama command
Qwen2.5 (3B)~2.0 GB8GB RAMBest for Korean, Chinese and Japanese. Keeps East Asian vocabulary well and translates tags very quickly.ollama run qwen2.5:3b
Llama 3.2 (3B)~2.0 GB8GB RAMBest for English and European languages. Ideal for tags in Spanish, Portuguese and other Western languages.ollama run llama3.2:3b
Gemma 2 (2B)~1.6 GB8GB RAMUltra-light for low-spec machines. Uses little memory, so it's easy to run alongside a vision model.ollama run gemma2:2b

3. Connect mysnipbox

  1. 1Make sure Ollama is running (the menu-bar/tray icon, or ollama serve).
  2. 2In mysnipbox, open Settings and choose Ollama as the provider.
  3. 3Keep the default endpoint http://localhost:11434, then select your vision model and your text model.
  4. 4Run a test image. The first request is slower while the model loads into memory.

Connect to a remote Ollama server

Ollama doesn't have to run on the same computer. Use a powerful PC on your home network, a home server, or a cloud GPU (for example RunPod) instead.

  1. 1In mysnipbox, open Settings and choose Ollama as the provider.
  2. 2Replace the Base URL field's http://localhost:11434 with your server's address: http://<server-IP>:11434.
  3. 3On the server, allow outside connections before you start Ollama (see the variables below), then pull the models there.
shell (server)
OLLAMA_HOST=0.0.0.0 OLLAMA_ORIGINS="*" ollama serve

OLLAMA_HOST=0.0.0.0 makes Ollama listen on every network interface, and OLLAMA_ORIGINS="*" lets requests from any origin through. If Ollama runs as a background app or service, set these in that service's environment and restart it.

Your own GPU server: Custom (OpenAI-compatible)

If you run your own GPU server with vLLM, LMDeploy, LocalAI or any engine that speaks the OpenAI API, connect it through the Custom provider.

  1. 1In mysnipbox, open Settings and choose Custom (OpenAI-compatible) as the provider.
  2. 2Fill in the three fields below.
  3. 3Run a test image to confirm the server answers.
FieldWhat to enter
Base URLYour server's OpenAI-compatible address, for example http://<GPU-IP>:8000/v1.
API KeyEnter it only if your server requires authentication. Servers on your own machine or private network without auth can leave this empty.
Model NameThe exact name of the vision model your server is serving, typed in by hand, for example Qwen/Qwen2.5-VL-7B-Instruct.

Troubleshooting

  • Connection refused: Ollama isn't running. Start it, or run ollama serve.
  • Model not found: pull it first with ollama pull <model>; the name must match exactly.
  • Very slow or crashes: the model is too big for your memory. Pick a smaller one, close other heavy apps, or use a cloud provider.
  • ollama ps shows which models are loaded and how much memory they use.