TopolearnFine-tuned and niche models often aren't in the library; importing lets you run them like any other model.
Not every model is in the Ollama library. You may have fine-tuned one yourself or found it on Hugging Face (a large public hub for models). Ollama can import two common file formats:
- GGUF: a single-file format from the llama.cpp project, usually already quantized (file names contain things like
Q4_K_M). - Safetensors: the standard format for original, unquantized weights, typically a folder with several files plus configuration.
Import with a Modelfile whose FROM points at the files:
FROM ./my-model-Q4_K_M.ggufor FROM /path/to/safetensors/folder, then run ollama create my-model. Split GGUF files keep their original names and can be referenced with a wildcard such as FROM ./model-*.gguf.
Quantization during import: for Safetensors imports, ollama create my-model --quantize int4 (short -q) creates a smaller quantized version. Current versions accept levels such as int4, int8, nvfp4, mxfp4 and mxfp8; an unsupported value prints the list for your version. GGUF files are not re-quantized; quantize them beforehand with llama.cpp's tools.
Prompt template: chat models are trained with a specific format for roles. Ollama detects common ones automatically. If an imported model gives strange, rambling answers, the template is likely wrong; copy the TEMPLATE from a library model of the same family (ollama show family --modelfile).
Sharing: create an account on ollama.com, add your Ollama public key in its settings, name the model with your username (ollama cp my-model myuser/my-model), then ollama push myuser/my-model. Others run it with ollama run myuser/my-model.
Respect licenses: many open-weight models restrict commercial use or redistribution.
Example
# 1) Import a GGUF file (already quantized)
cat > Modelfile <<'EOF'
FROM ./support-bot-Q4_K_M.gguf
PARAMETER temperature 0.3
SYSTEM You are Acme's support assistant.
EOF
ollama create support-bot -f Modelfile
ollama run support-bot "Where is my order?"
# 2) Import Safetensors weights and quantize while importing
echo "FROM ./my-finetune-safetensors" > Modelfile.st
ollama create --help # check supported --quantize levels
ollama create my-finetune -f Modelfile.st --quantize int4
# 3) Share on ollama.com (after adding your public key in account settings)
ollama cp support-bot myuser/support-bot
ollama push myuser/support-botRemember
FROM ./file.gguforFROM ./safetensors-folder, thenollama create--quantizeworks for Safetensors imports; GGUF must be quantized beforehand- Wrong TEMPLATE causes odd answers;
ollama push user/modelshares it
Watch out for
- Importing a chat fine-tune whose prompt template Ollama doesn't recognise and blaming the fine-tune for gibberish answers.