How to Run Qwen 3 Locally — Step-by-Step (2026)

Qwen3 (from Alibaba) is one of the best open model families you can run locally in 2026 — strong at reasoning, coding, and multilingual tasks. It comes in an unusually wide range of sizes, including two mixture-of-experts (MoE) models that punch above their weight. This guide picks the right 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

Every Qwen3 tag and its download size, from the official Ollama library:

Command	Download size	Type	Context	Comfortable VRAM
ollama run qwen3:0.6b	523MB	Dense	40K	CPU / any GPU
ollama run qwen3:1.7b	1.4GB	Dense	40K	4GB+
ollama run qwen3:4b	2.5GB	Dense	256K	6GB+
ollama run qwen3:8b	5.2GB	Dense	40K	8GB
ollama run qwen3:14b	9.3GB	Dense	40K	12–16GB
ollama run qwen3:30b	19GB	MoE (A3B)	256K	24GB
ollama run qwen3:32b	20GB	Dense	40K	24GB
ollama run qwen3:235b	142GB	MoE	256K	Server / multi-GPU

All sizes default to 4-bit (Q4_K_M) quantization. For the memory math behind these, see the VRAM requirements guide or the VRAM calculator.

The mapping most people need:

• 8GB GPU: qwen3:8b — strong all-rounder.
• 12–16GB GPU: qwen3:14b — the quality sweet spot.
• 24GB GPU: qwen3:30b (MoE) — the smart pick (see below), or qwen3:32b dense.

Why the 30B-A3B MoE Is the Standout

qwen3:30b is a mixture-of-experts model: it has 30B total parameters but only activates about 3B per token (hence “A3B”). In practice that means it needs the memory of a 30B model — about 19GB, so a 24GB card — but generates at closer to the speed of a much smaller model, because only a fraction of the weights run each step. It also ships with a 256K context window. If you have 24GB, it is often the best balance of quality and speed in the whole lineup. The dense qwen3:32b is also excellent but runs slower for similar memory.

Step-By-Step Setup

1. Confirm Ollama works

ollama --version

If that errors, install Ollama using our install guide.

2. Pull and run your size

For a 16GB card:

ollama run qwen3:14b

The first run downloads the model, then drops you into an interactive prompt. Qwen3 is strong at coding and structured reasoning, so test it with a real task rather than a trivia question.

3. Confirm it loaded on the GPU

ollama ps

If it shows GPU, you are at full speed. If it shows CPU or a split, the model did not fully fit — drop to the next smaller size.

4. Use the long-context sizes when you need them

The qwen3:4b, qwen3:30b, and qwen3:235b tags ship with a 256K context window — useful for long documents or large codebases. Remember that using long context costs extra VRAM (roughly +8–16GB at very long context), so a model that fits at short context can spill when you fill the window.

Common Mistakes

• Skipping the 30B MoE on a 24GB card. If you have 24GB, qwen3:30b (A3B) usually beats the dense qwen3:32b on speed for similar quality. Try it first.
• Pulling qwen3:235b on a desktop. It is a 142GB MoE — server/multi-GPU territory, not a consumer card.
• Maxing context without budgeting VRAM. The 256K-context sizes are great, but filling that window adds a lot of memory. Watch ollama ps.
• Picking a size that doesn’t fit. A 32B model on a 16GB card spills to system RAM and crawls. Match the tag to the table above.

What To Do Next

• Compare against a reasoning model: how to run DeepSeek R1 locally.
• Add a friendlier interface: LM Studio vs Ollama to decide between a GUI and the CLI.
• If 14B feels tight, the upgrade target is a 24GB card so the 30B MoE becomes comfortable — see the used RTX 3090 guide.
• Understand the 4-bit default: quantization explained.

Frequently Asked Questions

Which Qwen3 size should I run?

Match it to VRAM: qwen3:8b for 8GB, qwen3:14b for 12–16GB, and qwen3:30b (the efficient MoE) for a 24GB card. The 14B dense model is the best balance for most mid-range GPUs.

What is Qwen3-30B-A3B?

It is a 30B mixture-of-experts model that activates only about 3B parameters per token. It needs ~19GB of VRAM (a 24GB card) but runs much faster than a dense 30B model because only a fraction of the weights are used each step. It also has a 256K context window.

How much VRAM does Qwen3 need?

It depends on size: ~5GB for the 8B, ~9GB for the 14B, and ~19–20GB for the 30B/32B (a 24GB card). The tiny 0.6B–4B models run on minimal hardware or CPU.

Is Qwen3 good for coding?

Yes — Qwen3 is one of the stronger open families for code and structured reasoning. The 14B and 30B sizes are good local coding assistants; pick based on your VRAM.

Can I run Qwen3 with a long context?

Yes. The 4B, 30B, and 235B tags ship with a 256K context window. Just budget extra VRAM for long context — filling the window can add many GB, so a model that fits at short context may spill when fully loaded.

---

Last updated: May 2026. Model tags, download sizes, context lengths, and dense/MoE designations reference the official Ollama Qwen3 library at time of publication. VRAM targets reference our VRAM requirements guide.