Free Ollama Models: 10 Best to Run Locally (No API Key)

Published November 6, 2025 · Updated November 2, 2026 · 12 min read

Top pick: ollama pull qwen3:8b (5.2GB). Every model in the Ollama library that you download is free: no account, no API key, no subscription, and it runs offline once pulled. For an 8GB machine the best free picks are Qwen3 8B (general use, tool calling, a thinking mode), Gemma 4 E4B (also reads images and audio), Llama 3.1 8B (the most-pulled model in the library) and Phi-4 Mini when you only have 4GB. The one thing that is not free is the cloud tags — anything ending in :cloud or -cloud runs on Ollama's servers and is billed per token (section below). Which local model is right for you is settled by how much memory you have, and the sizes below are the real download sizes.

Which free Ollama models should you install first?

Ten free models, the exact command to install each, and the download size the Ollama registry lists for that tag (checked November 2, 2026). Nothing here costs money, and nothing here needs an internet connection once the download finishes.

All ten are open-weight and free to run. Licences differ (Apache 2.0 for the Qwen models, Gemma 4 and gpt-oss; MIT for Phi-4 Mini and DeepSeek-R1; Meta's community licence for Llama) — read the licence on the model page before you build a product on one. For the task-by-task ranking of the whole library, including the 24 GB and 48 GB picks, see our best Ollama models guide; this page is about what you can run for free on the machine you already have.

If you are choosing blind: install Qwen3 8B. It fits in 8 GB, does tool calling for agents and apps, and has a thinking mode you can switch off for quick answers. If you want the model with the most tutorials written for it, Llama 3.1 8B is the safe alternative.

Which Ollama models are free, and which are cloud-only?

The Ollama library now lists two kinds of model, and only one of them is free:

Ollama's cloud docs say "Cloud models do not need to be downloaded" and that you sign in to use them; its pricing page lists the plans (read November 2, 2026). If you want to be sure nothing ever leaves your machine, the Ollama FAQ documents a local-only switch: set OLLAMA_NO_CLOUD=1 (or "disable_ollama_cloud": true in ~/.ollama/server.json) and restart Ollama.

Ollama's cloud is not the only hosted option with a no-cost entry point. If you want a hosted API that costs nothing while you prototype, the Groq free tier serves open models such as Llama 3.1 8B and Llama 3.3 70B behind per-minute and per-day rate limits. Like any cloud tag, your prompts leave your machine.

Real download sizes and what fits

Sizes are the Ollama registry's figures for each tag at its default quantization (Q4_K_M for most). Add 1-2 GB for the context (KV cache) and leave room for your operating system.

---

Older free models that still work

The previous edition of this page recommended the ten models below. They are all still free and still in the library, and plenty of existing tutorials and prompts assume them, so here they are with real tags and download sizes. The newer picks above beat most of them at the same size.

---

How much RAM does each free model need?

You do not need a chart for this. Quantized weights are roughly linear in parameter count, and at Q4_K_M — the default quantization Ollama pulls for most tags — the rule of thumb is:

weights in GB  ≈  0.6 × (parameters in billions)

then add ~1-2 GB for the KV cache at a 4K-8K context,
plus whatever your operating system is already using.

Worked examples, so you can check the arithmetic (the real download sizes in the table above land close to this line; models that bundle a vision or audio encoder, like Gemma 4, run larger):

The "practical system RAM" column is the next common memory tier above weights + cache + OS. It is guidance, not a hard floor: a model can run in less if you shorten the context or drop to a harsher quant, and it will run badly if the weights spill to disk.

Heavier quants change the multiplier, not the method. Q8_0 is roughly 1.0 GB per billion parameters, Q5_K_M about 0.7, Q3_K_M about 0.5. Our quantization comparison covers what you give up at each step.

How fast will these models run on your hardware?

Generating one token requires reading every weight once, so throughput has a hard arithmetic ceiling:

tokens/sec ceiling  =  memory bandwidth (GB/s)  ÷  model size (GB)

That is an upper bound, not a prediction. Real output lands below it — prompt processing, sampling, context growth and background load all take time — but the ceiling tells you instantly whether a model is viable on a machine. If the ceiling is 6 tok/s, no amount of tuning makes that model feel interactive.

Where the bandwidth column comes from: the GPU and Apple rows are the vendors' published memory specifications. The DDR rows are computed from the JEDEC data rate, MT/s × 8 bytes per channel × number of channels — so DDR4-3200 in dual channel is 3200 × 8 × 2 = 51.2 GB/s, and DDR5-5600 in dual channel is 89.6 GB/s. Every tok/s cell is then just bandwidth ÷ model size. It is all arithmetic, not a benchmark run.

For a real-world check, our reference measured run — an Apple M3 Pro (150 GB/s), Ollama 0.40.13, 2 November 2026 — got 27.2 tok/s from llama3.1:8b (ceiling ~30), 25.3 from qwen3:8b, 43.7 from phi4-mini and 55.6 from llama3.2:3b: roughly 73-89% of each model's ceiling. On a rented NVIDIA L4 (24 GB, 300 GB/s) the same llama3.1:8b tag measured 49.2 tok/s (Ollama 0.40.13, 3 November 2026; every row is on /benchmarks).

Two things fall straight out of this table. First, GPU offload is worth far more than a faster CPU — the gap between system RAM and GDDR6X is roughly an order of magnitude. Second, on a CPU-only box the smaller model is not a compromise, it is the only interactive option. That is why Phi-3 Mini keeps its place in the list.

For the bandwidth number for your specific card, see our GPU memory bandwidth guide; if things are slower than the ceiling suggests they should be, why local LLMs run slow walks the usual causes.

What is each free model actually good at?

The current ten, in one line each:

• Qwen3 8B — Alibaba's Apache 2.0 all-rounder; carries the tools and thinking badges in the library, so it suits agents and apps as well as chat.
• Gemma 4 E4B — Google's laptop-sized Gemma 4: text, images and audio in one 6.6 GB download.
• Llama 3.1 8B — the most-pulled model in the library; every integration supports it.
• Llama 3.2 3B — Meta's 2 GB model for old laptops; 55.6 tok/s measured on our M3 Pro.
• Phi-4 Mini — Microsoft's MIT-licensed 3.8B reasoning model in 2.5 GB.
• Qwen 2.5 Coder 7B — the most-pulled coding family in the library, 4.7 GB.
• DeepSeek-R1 8B — shows its reasoning before answering; now built on Qwen3 8B.
• Gemma 4 12B — images plus text with a 256K context on a 12 GB card.
• gpt-oss 20B — OpenAI's open-weight model, Apache 2.0; a MoE that the model card says runs "on systems with as little as 16GB memory".
• Qwen3.8 27B — the strongest dense model that fits one 24 GB GPU; tools, thinking and vision.

The rest of this section covers the older ten, which are still worth knowing.

1. Llama 3.1 8B — the default choice

Install: ollama pull llama3.1:8b

Meta's Llama 3.1 8B is the most-pulled general-purpose open model, which matters more than it sounds: tooling, quantizations, fine-tunes and troubleshooting answers all exist for it first. Released under Meta's Llama 3.1 Community License, which permits commercial use below 700M monthly active users.

• RAM: 8 GB · Storage: 4.9 GB · GPU optional
• Good for: writing, summarizing, general Q&A, light coding
• Full breakdown in our 8 GB RAM model guide; if writing is the main job, see the best local AI models for writing

2. Mistral 7B v0.3 — smallest model that still feels general

Install: ollama pull mistral:7b (the library's default mistral tag is v0.3)

Mistral 7B is Apache 2.0 — the most permissive licence of anything in this list, with no user-count clause and no acceptable-use rider. v0.3 added a larger vocabulary and function-calling support over v0.2. Slightly smaller than Llama 3.1 8B, so its bandwidth ceiling is slightly higher on the same machine.

3. Phi-3 Mini — the 4 GB option

Install: ollama pull phi3:mini

Microsoft's Phi-3 Mini is 3.8B parameters under an MIT licence, trained on a heavily filtered "textbook quality" corpus rather than raw web scrape. At 2.2 GB quantized it is one of the few older models with a usable ceiling on DDR4 laptops and older desktops; Phi-4 Mini (2.5 GB) is its successor.

4. Gemma 2 9B — Google's open-weight model

Install: ollama pull gemma2:9b

Gemma 2 9B uses interleaved local/global attention and knowledge distillation from a larger teacher model. Distributed under Google's Gemma Terms of Use — free for commercial use, but with a use-restriction policy attached, so read it if you are shipping a product.

5. Qwen 2.5 7B — multilingual and code-capable

Install: ollama pull qwen2.5:7b

Alibaba's Qwen 2.5 7B is Apache 2.0 and documented for 29 languages. The Qwen family also ships coder and math variants at the same size, so you can swap specialization without changing your RAM budget. Newer generation: Qwen3 8B in the table above, and our Qwen 3 setup guide.

6. CodeLlama 13B — code-specialized

Install: ollama pull codellama:13b

CodeLlama 13B is Llama 2 further-trained on code, with infilling support that makes it usable for editor autocomplete rather than just chat. Needs 16 GB. Pair it with Continue.dev for an in-editor assistant.

7. DeepSeek Coder 6.7B — code on an 8 GB machine

Install: ollama pull deepseek-coder:6.7b

DeepSeek Coder 6.7B is trained on a code-heavy corpus with repository-level context, and its 33B sibling exists if you have 32 GB to spend on it. The 6.7B is the practical pick when CodeLlama 13B will not fit.

8. Solar 10.7B — depth up-scaled 10.7B

Install: ollama pull solar

Upstage's Solar 10.7B was built by depth up-scaling a 7B base — duplicating and re-training layers rather than training a larger model from scratch. Apache 2.0.

9. Vicuna 13B — conversational fine-tune

Install: ollama pull vicuna:13b

Vicuna 13B is LMSYS's Llama 2 fine-tune on shared conversation data. Older than everything else here; kept in the list because its chat style is distinctive and a lot of existing prompts assume it.

10. OpenHermes 2.5 — instruction and structured output

Install: ollama pull openhermes

OpenHermes 2.5 is a Mistral 7B fine-tune aimed at instruction following and structured (JSON-shaped) responses. Same RAM footprint as base Mistral.

For the ranked, task-by-task view across the whole Ollama library, see our best Ollama models roundup.

What do free local models save against paid AI subscriptions?

The comparison is only honest as arithmetic on published list prices. ChatGPT Plus and Claude Pro both list at $20/month; GitHub Copilot Pro lists at $10/month. Multiply by twelve:

Check the vendors' current pricing before quoting those figures anywhere — they move.

What that table does not include is electricity and hardware. A local model is free to licence, not free to run, and a frontier cloud model is still better at hard reasoning than an 8B running on your laptop. The honest case for local is privacy, offline availability, no rate limits and no per-token meter — see local AI vs ChatGPT on cost for the full comparison.

How do you install a free Ollama model?

1. Install Ollama

# macOS / Linux
curl -fsSL https://ollama.com/install.sh | sh

# Windows: download the installer from ollama.com

2. Pull a model

ollama pull qwen3:8b

3. Run it

ollama run qwen3:8b

On macOS and Windows you can do the same from Ollama's own desktop app instead of the terminal.

That is the whole flow. Windows users who want the step-by-step version have our Ollama Windows installation guide; if you would rather have a browser chat interface than a terminal, set up Open WebUI.

Housekeeping worth knowing on day one:

• ollama list shows what you have downloaded and how much disk it uses
• ollama rm <model> deletes one — models are large, this matters
• ollama ps shows what is currently loaded in memory
• Set OLLAMA_KEEP_ALIVE to control how long a model stays resident after a request

How do you pick a model for your machine?

Start with your RAM, not with a ranking. Run the 0.6 × parameters calculation above against what you actually have free, and the shortlist writes itself.

• 4 GB RAM: Llama 3.2 3B or Phi-4 Mini (Phi-3 Mini still works too).
• 8 GB RAM: Qwen3 8B or Llama 3.1 8B for general use, Qwen 2.5 Coder 7B for code, DeepSeek-R1 8B for reasoning.
• 16 GB RAM / 12 GB GPU: Gemma 4 E4B or Gemma 4 12B for images and long documents; gpt-oss 20B on a 16 GB GPU.
• 32 GB RAM or a 24 GB GPU: Qwen3.8 27B, plus the 30B-class models in our best Ollama models guide. See what a 24 GB card actually holds in our RTX 3090 local AI guide.

Second question: CPU or GPU. Every model here runs on CPU, and the bandwidth table shows exactly what that costs you. If you are on CPU only, stay at 8B or below and expect single-digit to low-double-digit tokens per second.

Frequently asked questions

Are these models really free?

Yes, every model you download from the Ollama library is free to download and run, including all twenty on this page. "Free" is not the same as "unrestricted", though — Mistral 7B, Qwen 2.5 and Solar are Apache 2.0 with no strings; Phi-3 Mini is MIT; Llama 3.1 and CodeLlama carry Meta's Community License with a 700M-monthly-active-user threshold; Gemma 2 carries Google's use policy. Read the licence text on the model's Hugging Face card before you build a product on one.

Are Ollama cloud models free?

No. Tags ending in :cloud or -cloud run on Ollama's servers, need a sign-in, and are billed per token; Ollama's pricing page lists a $0 "Pay as you go" plan that pays for each request with usage credits, and a $20/month Pro plan. To make sure you only ever use free local models, set OLLAMA_NO_CLOUD=1 and restart Ollama.

How do these compare to ChatGPT?

An 8B model running on your laptop is not a frontier model and will not match one on hard reasoning, long-context work or breadth of world knowledge. What it does match is availability: no rate limits, no per-token cost, no outage, and nothing sent to a third party. For drafting, summarizing, reformatting, code completion and Q&A over your own files, that trade is usually worth it.

Do I need a GPU?

No. All ten run on CPU. The throughput formula above tells you what to expect: divide your memory bandwidth by the model size. A GPU raises the ceiling roughly tenfold because GDDR6X bandwidth is roughly ten times dual-channel DDR5 — it is not a small difference, but it is not a requirement either.

Can I run several models at once?

You can install as many as your disk allows and switch instantly. Running them simultaneously is limited by memory — each loaded model holds its full weights resident. Use ollama ps to see what is loaded and OLLAMA_KEEP_ALIVE to unload idle models.

Which quantization should I download?

Q4_K_M unless you have a reason not to. It is Ollama's default for most tags and the best quality-per-gigabyte point in common use. Go to Q5_K_M or Q8_0 if you have memory to spare and want fewer errors; drop to Q3 only when the alternative is not running the model at all.

Your next steps

1. Browse the best Ollama models — the whole library ranked by task
2. Set up Open WebUI — a browser chat interface for your models
3. Install your first local AI — complete beginner walkthrough
4. Understand quantization — why models come in different sizes
5. Set up AI coding — a free in-editor assistant
6. Try GPT-OSS — OpenAI's open-weight release
7. Run Llama 4 Scout — Meta's MoE with a very long context window
8. Run Qwen3-Coder locally — the coding-specialized Qwen
9. Set up Qwen 3 locally — full install for Alibaba's newer flagship