How to Run Llama Locally — Step-by-Step (2026)
Run Meta's Llama models on your own GPU with Ollama. Which size fits your VRAM (3.2, 3.1, 3.3), exact commands, and how to get the best results locally.
Prerequisites
- ✓Ollama installed
- ✓A GPU with at least 8GB VRAM for a usable experience
- ✓Disk space matching your chosen model (1.3–243GB)
Quick answer
Install Ollama, then run ollama run llama3.1:8b for an 8GB GPU — the best all-round local Llama. On a 24GB card stick with 8B (Llama has no mid-size dense option between 8B and 70B), and run llama3.3:70b only across dual 24GB GPUs or a big unified-memory Mac. For tiny hardware, llama3.2:3b or llama3.2:1b. Llama is the most-downloaded open model family, and 8B is the workhorse most people should start with.
Meta’s Llama is the most-downloaded open model family for local use, and for good reason: the 8B model is a fast, reliable general-purpose workhorse that fits an 8GB GPU. This guide picks the right Llama size for your hardware and runs it in one command.
If Ollama is not installed yet, start with our Ollama install guide.
Pick the Right Size for Your VRAM
The Llama tags and download sizes from the official Ollama library:
| Command | Download size | Context | Comfortable VRAM |
|---|---|---|---|
ollama run llama3.2:1b | 1.3GB | 128K | CPU / any GPU |
ollama run llama3.2:3b | 2.0GB | 128K | 4GB+ |
ollama run llama3.1:8b | 4.9GB | 128K | 8GB |
ollama run llama3.3:70b | 43GB | 128K | 48GB (dual 24GB) |
ollama run llama3.1:405b | 243GB | 128K | Server / many GPUs |
All sizes default to 4-bit (Q4_K_M) quantization and ship with a 128K context window. For the memory math, see the VRAM requirements guide or the VRAM calculator.
The mapping most people need:
- 8GB GPU or Apple Silicon:
llama3.1:8b— the default workhorse. - 24GB GPU: still
llama3.1:8bfor Llama specifically. Llama has no dense model between 8B and 70B, so on 24GB you either run 8B fast or look to other families (Qwen3 30B, DeepSeek R1 32B) for the mid-size tier. See best local LLM models 2026. - Dual 24GB / big Mac:
llama3.3:70b— the best general open model at the large tier. - Tiny hardware / edge:
llama3.2:3bor1b.
A Note on Llama Versions
The Llama family spans a few generations, and the version matters:
- Llama 3.1 — the workhorse 8B (plus 70B and the giant 405B).
- Llama 3.2 — small, efficient 1B and 3B models for edge and low-power devices.
- Llama 3.3 — a 70B model that delivers the family’s best general quality at the large tier.
For a single 8GB–16GB card, llama3.1:8b is the one to run. The 70B models are a different hardware class.
Step-By-Step Setup
1. Confirm Ollama works
ollama --versionIf that errors, install Ollama using our install guide.
2. Pull and run
ollama run llama3.1:8bThe first run downloads the model (about 5GB), then drops you into an interactive chat. Llama 3.1 8B is a strong generalist — good for chat, writing, summarizing, and light coding.
3. Confirm it loaded on the GPU
ollama psIf it shows GPU, you are at full speed. If it shows CPU or a split, the model did not fully fit — though 8B fits almost any 8GB+ card comfortably.
4. Use the long context when you need it
Every listed Llama size ships with a 128K context window — useful for long documents. Remember that filling a long context costs extra VRAM, so a model that fits at short context can get tight when the window is full.
Common Mistakes
- Expecting a mid-size dense Llama. There is no dense Llama between 8B and 70B. If you have a 24GB card and want a ~30B model, look at Qwen3 30B or DeepSeek R1 32B instead.
- Trying llama3.1:405b on a desktop. It is a 243GB model — server/multi-GPU territory, not a consumer card.
- Running 70B on a single 24GB card and expecting comfort. Llama 3.3 70B needs ~38GB+ at Q4 — practical across dual 24GB GPUs with runtime-supported sharding, or on a large unified-memory Mac.
- Picking 1B/3B for serious work. The Llama 3.2 small models are for edge and speed, not quality. For real tasks, 8B is the floor.
What To Do Next
- Compare Llama with other families: best local LLM models 2026.
- Try a reasoning model: how to run DeepSeek R1 locally, or a strong coder: how to run Qwen 3 locally.
- Add a ChatGPT-style interface: how to set up Open WebUI.
- Want 70B comfortably? You need roughly 48GB of usable aggregate capacity — see the used RTX 3090 guide for runtime-sharded dual GPUs, or use a large unified-memory system.
Frequently Asked Questions
Which Llama model should I run locally?
For most people, llama3.1:8b — it fits an 8GB GPU and is a strong generalist. Use llama3.2:3b/1b for edge devices, and llama3.3:70b only on dual 24GB GPUs or a large unified-memory Mac.
How much VRAM does Llama 3.1 8B need?
About 8GB of VRAM at 4-bit, with room for context — it runs comfortably on any 8GB+ card or Apple Silicon. The download is roughly 5GB.
Can I run Llama 3.3 70B locally?
Yes, but it needs ~38GB+ of VRAM at Q4 — practical across two 24GB GPUs only when the runtime shards the model, or on a large unified-memory Mac. NVLink can improve transfers between RTX 3090s but does not merge their VRAM.
Is there a mid-size Llama between 8B and 70B?
Not as a dense model. If you have a 24GB card and want a ~30B model, look to other families — Qwen3 30B (an efficient MoE) or DeepSeek R1 32B. See best local LLM models 2026.
What’s the difference between Llama 3.1, 3.2, and 3.3?
Llama 3.1 has the workhorse 8B (plus 70B and 405B); Llama 3.2 adds small 1B/3B edge models; Llama 3.3 is a 70B with the family’s best general quality. For a single GPU, llama3.1:8b is the one to run.
Last updated: June 2026. Model tags, download sizes, and context lengths reference the official Ollama Llama library at time of publication. VRAM targets reference our VRAM requirements guide.