Key takeaways
- The Ollama app is small next to its models; budget disk space for models, from about 1 GB to tens of GB each.
- A model's size is roughly parameters times bytes per parameter: about 0.5 to 0.6 GB per billion parameters at 4-bit quantisation.
- Under 5 GB, 3B to 8B models at 4-bit run on most laptops and handle everyday tasks well.
- Ollama's default context window is modest; raise num_ctx for long documents, at the cost of memory.
- On Windows, the native app is simplest; use WSL when your development already lives there.
Ollama itself is a small download compared with the models it runs; the space goes to models, which range from about 1 GB for tiny ones to 4 to 5 GB for popular 7B and 8B models at 4-bit, and tens of gigabytes for large ones. This page explains how to estimate a local model's size, which models fit under 5 GB, what parameters and tokens mean in Ollama, how many models it supports, where it installs on Windows, whether to use Windows or WSL, and how to get started.
How many GB is Ollama?
The Ollama application is small relative to models, though its Windows installer includes GPU libraries that add some size. Plan your disk space around the models: a typical setup with three or four models uses 10 to 25 GB.
Check sizes with ollama list, which shows each installed model and its size on disk, and free space with ollama rm <model>. The download page shows the current installer size. If your system drive is small, store models elsewhere with the OLLAMA_MODELS environment variable. Remember that running a model also needs RAM or video memory roughly equal to its file size plus extra for the context, so disk space is not the only limit.
How many GB is a local LLM?
Roughly the number of parameters multiplied by the bytes stored per parameter. At 16-bit precision that is 2 bytes per parameter; at the 4-bit quantisation common in Ollama it is about 0.5 to 0.6 bytes, so a 7B or 8B model is around 4 to 5 GB.
Approximate sizes at 4-bit:
- 1B to 3B models: about 1 to 2 GB.
- 7B to 8B models: about 4 to 5 GB.
- 12B to 14B models: about 8 to 9 GB.
- 30B to 34B models: about 18 to 20 GB.
- 70B models: about 40 GB.
Higher-precision versions, such as 8-bit, are about twice as large and slightly more accurate. Quantisation is the same idea QLoRA uses to fine-tune large models on one GPU: store weights in 4 bits to save memory.
What is the best Ollama model under 5 GB?
For general use, a 7B or 8B instruction model at 4-bit, from families such as Llama, Qwen or Mistral, gives the best quality under 5 GB. For low-memory laptops, 3B to 4B models such as small Llama, Gemma or Phi variants are faster and still useful.
Also asked as: best ollama model under 5gb
Choose by task:
- Chat and writing: a recent 7B or 8B instruct model.
- Coding: a coding-tuned 7B model, such as a Qwen coder variant.
- Speed on 8 GB RAM: a 3B or 4B model.
- Embeddings for RAG: a dedicated embedding model, which is much smaller.
New models arrive often and smaller ones keep improving, so check the Ollama library and compare two candidates on your own prompts. Quality differences show quickly on real tasks.
How many parameters do Ollama models have?
Ollama runs models of many sizes, from under one billion parameters to hundreds of billions. The parameter count is usually in the model tag, such as llama3.2:3b or qwen2.5:7b, and ollama show <model> displays details.
Also asked as: how many parameters in ollama
Parameters are the learned weights of the network. More parameters generally mean more knowledge and better reasoning, but also more memory and slower responses. What you can run depends on your RAM or GPU memory: roughly, a 4-bit model needs a little more memory than its file size. On a laptop with 16 GB RAM, models up to about 8B to 14B are practical; beyond that you need a strong GPU or a server.
How many models does Ollama support?
Ollama's library offers hundreds of models and many more variants, covering major open families such as Llama, Qwen, Gemma, Mistral, Phi and DeepSeek, plus embedding and vision models. You can also import other compatible models in GGUF format.
To import a model from Hugging Face or your own fine-tune, create a Modelfile pointing to the GGUF file and run ollama create. Many GGUF repositories on Hugging Face can also be run directly by reference. Each model has tags for sizes and quantisation levels, so one family may have dozens of versions. Stick to well-known publishers and popular variants unless you have a reason to experiment, because quality varies widely among community uploads.
How many tokens can Ollama handle?
It depends on the model's maximum context length and Ollama's num_ctx setting. Ollama uses a modest default context window, often a few thousand tokens, even when the model supports far more, to keep memory use reasonable.
Also asked as: how many tokens ollama
Raise it per request through the API options, in a Modelfile with PARAMETER num_ctx 16384, or with environment settings in recent versions. A larger context lets you pass longer documents or conversations, but uses more memory and slows responses. If answers ignore the start of a long document, the context window is likely too small and the text is being cut off. For long documents, retrieval of relevant chunks usually works better than one huge context, especially with small models.
Where is Ollama installed on Windows?
By default, Ollama installs into your user profile's local app data folder, and models are stored in C:\Users\<you>\.ollama\models. The ollama command is added to your path, and the app runs from the system tray.
Also asked as: how to ollama windows
To store models on another drive, set the OLLAMA_MODELS environment variable and restart Ollama. Logs, useful when something fails, are in the Ollama folder under your local app data; the tray icon's menu can open them. Ollama's local API listens on http://localhost:11434. If another app already uses that port, or you want to reach Ollama from other devices, set OLLAMA_HOST, and add protection before exposing it beyond your machine.
Should I run Ollama on Windows or WSL?
Run it natively on Windows unless your development work already lives in WSL. The Windows app is simpler, starts automatically and uses supported GPUs directly; WSL makes sense when your Python projects, Docker setup and tools are all inside Linux.
Either way, you can reach Ollama from the other side: WSL can call the Windows app over the local network, and Windows tools can reach Ollama in WSL through forwarded localhost ports. Avoid running two copies on the same port. WSL needs enough memory assigned in its settings, or larger models fail to load. If GPU acceleration does not work in WSL, update the Windows GPU driver first; the WSL GPU support depends on it.
Is there an Ollama tutorial for Windows?
Yes. Install from ollama.com, open PowerShell, run a model, then connect it to a chat interface or your code. A first session takes about fifteen minutes.
Also asked as: tutorial ollama windows
ollama run llama3.2 # downloads, then opens a chat
ollama list # installed models
ollama ps # running models and CPU or GPU use
Then try the API from Python:
import requests
r = requests.post("http://localhost:11434/api/generate",
json={"model": "llama3.2", "prompt": "Explain BM25 briefly.", "stream": False})
print(r.json()["response"])
For a graphical chat, use Ollama's own app interface or install a web UI that connects to the local API.
How do I use Ollama on Windows day to day?
Keep Ollama running in the tray, pull the few models you actually use, and work from either the chat interface, the terminal or your code through the local API. Clean up models you no longer need.
Useful habits:
ollama pull <model>to update a model to its latest version.ollama stop <model>to free memory after use.ollama rm <model>to reclaim disk space.- Use a coding editor extension that supports Ollama for local code help.
- Point LangChain or your own scripts at
http://localhost:11434.
If responses are slow, check ollama ps to see whether the model runs on the GPU; if it shows CPU, the model may be too large for your video memory, so try a smaller size or quantisation.
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.
