Key takeaways
- Get models from the Ollama library with ollama pull, or run compatible GGUF models from Hugging Face.
- Models are stored in ~/.ollama/models on Mac and Windows user installs, and under the ollama service user on Linux.
- Change the location with the OLLAMA_MODELS environment variable, then restart Ollama.
- Default model tags are usually 4-bit quantized; other tags offer higher precision.
- Local models keep data on your machine; Ollama unloads idle models after about five minutes by default.
Get Ollama models from the official Ollama library with ollama pull <model>, or run compatible GGUF models directly from Hugging Face. Ollama stores downloaded models in a .ollama/models folder in your home directory on Mac and Windows, and under the Ollama service account on Linux, and you can move them with the OLLAMA_MODELS environment variable. I run local models for private document work, so this page covers getting and storing models, how storage works, quantization, privacy, unloading, training and download blocks.
Where do I get Ollama models?
The main source is the Ollama library at ollama.com, which lists models such as Llama, Qwen, Gemma, Mistral, DeepSeek and embedding models, each with size and quantization tags. Download with ollama pull llama3.2 or run directly with ollama run.
Also asked as: where to get ollama models
You can also run many GGUF models hosted on Hugging Face with ollama run hf.co/<user>/<repo>, choosing a quantization tag, or import your own GGUF files with a Modelfile. Check each model's licence and size before downloading, since larger models need more memory and disk space. Ollama also offers cloud-hosted models for very large models that do not fit locally. For fine-tuned models of your own, export them to GGUF and create an Ollama model with a Modelfile, which is how a QLoRA model from FineTune Studio-style training becomes a local model.
Where does Ollama store models?
By default, Ollama stores models in ~/.ollama/models on macOS, C:\Users\<username>\.ollama\models on Windows, and, for the Linux system service install, under the ollama user's home, typically /usr/share/ollama/.ollama/models.
Inside, models are split into blobs, the large weight files, and manifests describing each model. Use ollama list to see installed models and their sizes, and ollama rm <model> to delete ones you no longer need. Models can take several gigabytes each, so monitor disk space, especially on laptops with small SSDs. Shared layers between tags of the same model are stored once, so pulling a second tag may use less space than expected.
Where does Ollama store models on a Mac?
On macOS, Ollama stores models in ~/.ollama/models, inside your user's home folder. The folder is hidden by default; open it in Finder with Go, Go to Folder, and type ~/.ollama/models, or browse it from Terminal.
Also asked as: where ollama store models mac
Apple silicon Macs run Ollama efficiently using unified memory, so larger models are practical on Macs with more RAM. To save internal disk space, move the models folder to an external SSD by setting OLLAMA_MODELS to the new path and restarting Ollama. Do not edit or rename files inside the blobs folder manually.
How do I change where Ollama stores models?
Set the OLLAMA_MODELS environment variable to your preferred folder and restart Ollama. On Windows, add it in System Environment Variables, then quit and restart Ollama; on macOS, set it for the Ollama app; on Linux, add it to the systemd service configuration.
Also asked as: where to change ollama models
You can move existing models by copying the models folder to the new location before setting the variable, or simply pull models again. Ensure the folder has enough space and that the user running Ollama has permission to read and write it. Check with ollama list that models appear after restarting.
How does Ollama store models?
Ollama uses a content-addressed storage format similar to container images: model weights and other layers are stored as blobs named by their SHA-256 hashes, and manifests list which blobs make up each model and tag, including parameters and templates.
This design deduplicates shared layers between models and tags, verifies downloads by hash, and makes updates efficient. The weights themselves are GGUF-format files, the format used by llama.cpp. Because files are referenced by hash, renaming or moving individual blobs breaks models; manage them with Ollama commands instead.
How many Ollama models are there?
The Ollama library lists hundreds of model families, each with many size and quantization tags, and thousands more compatible GGUF models are available on Hugging Face. New models are added frequently as open-weight releases appear.
Library models cover general chat, coding, reasoning, vision, embeddings and tool use. Rather than counting, choose by task and hardware: small models of one to four billion parameters for laptops, seven to fourteen billion for decent GPUs, and larger models for high-memory machines or Ollama's cloud models. Read model cards for licences, context length and capabilities.
Are Ollama models quantized?
Usually, yes. The default tags for most Ollama library models are quantized, commonly to around 4 bits using formats such as Q4_K_M, balancing size, speed and quality. Other tags offer 8-bit or 16-bit versions for higher fidelity.
Quantization lets large models run on consumer hardware: a 7 to 8 billion parameter model needs roughly 4 to 5 GB at 4 bits versus about 15 GB in 16-bit. Check available tags on the model's library page and pull a specific tag, such as :q8_0, if quality matters more than memory. Compare outputs on your tasks.
Are Ollama models private?
When you run local models with Ollama, prompts and data are processed on your own machine and are not sent to a cloud service, which makes Ollama popular for private documents and offline work. Ollama's cloud models, if you choose them, run on external servers.
Privacy also depends on your setup: if you expose Ollama's API on a network, anyone with access can use it, so keep it bound to localhost or protected. Applications built on top of Ollama may log data, so review their behaviour. For sensitive work, local models plus local storage, such as a local vector store for RAG, keep data fully on-device.
When does Ollama unload models?
By default, Ollama keeps a model loaded in memory for about five minutes after the last request, then unloads it to free memory. You can change this with the keep_alive request parameter or the OLLAMA_KEEP_ALIVE environment variable.
Set a longer keep-alive to avoid reload delays for frequently used models, or zero to unload immediately after each request on memory-constrained machines. Run ollama ps to see loaded models and ollama stop <model> to unload one manually. Loading multiple models at once requires enough RAM or VRAM; Ollama manages how many stay loaded.
Why does Ollama get blocked?
Ollama downloads can be blocked by corporate or college firewalls and proxies, antivirus or endpoint security software, network filters that block large downloads or specific domains, or regional network issues. The local API can also be blocked by firewall rules.
Also asked as: why ollama get blocked
Fixes: try a different network, configure the HTTPS_PROXY variable if your network requires a proxy, ask IT to allow Ollama's download domains, add an antivirus exception if you trust the installation, and check that port 11434 is allowed locally. On managed work devices, follow your organisation's policies before installing AI software.
Sources
GenAI engineer and AI product builder with 2+ years shipping production-grade AI systems: RAG pipelines, fine-tuned LLMs, hybrid retrieval and multi-modal apps across vision, speech and OCR, architected end to end from ingestion to deployment. Leads engineering for SCULT INDIA's 14-member team, founded the 500+ member TechVerse Enclave community and has mentored 200+ students. Three hackathon first prizes: Changethon 2025 (IIT Roorkee), Product Genesis at Vividhotsava 2025 (CSJMU) and BYTEBATTLE (MeetKats).
Open to GenAI roles, hackathon judging, mentorship sessions and guest talks at colleges. For campus sessions the inviting institution covers the visit end to end: travel (a First AC train ticket for journeys up to five hours, a flight for anything longer), accommodation, meals, and pickup and drop; the honorarium and workshop fees are negotiable on WhatsApp or on the call, and requests are welcome from institutions that can cover the visit this way. Email pranjulrathour41@gmail.com.
