← all posts
[ guide ]June 24, 202610 min read

How to Run Qwen Locally (2026 Guide)

Alibaba's Qwen3 family — including the standout Qwen3-Coder models — is some of the best open-weights you can run at home. A 2026 guide to picking a size, running it on your GPU, and making it reachable from your editor anywhere.

Alibaba's Qwen3 family is, in 2026, some of the most capable open-weights you can run on your own hardware — and the Qwen3-Coder models in particular have become a default choice for people running local coding agents. This guide covers the part that matters: picking a Qwen size that fits your GPU (including the clever mixture-of-experts option), getting it running, and wiring it into your editor from anywhere.

Step 1 — Pick a Qwen size (and notice the MoE trick)

Qwen3 ships in an unusually wide range, from sub-1B up to a 235B mixture-of-experts flagship. The ones that matter for local use:

ModelParams~VRAM (Q4_K_M)Good for
Qwen3 4B4B~3 GBLaptops, fast tasks
Qwen3 8B8B~5–6 GBThe general-purpose default
Qwen3 14B14B~10 GBBetter reasoning, 12 GB+ GPUs
Qwen3 32B32B~20 GBFrontier-ish, 24 GB card
Qwen3-30B-A3B (MoE)30B / ~3B active~18 GBBig quality, small-model speed
Qwen3-Coder-30B30B~18 GBBest local coding agent model

The standout is Qwen3-30B-A3B: 30B total parameters but only ~3B activate per token. Since local speed is bound by memory read per token, it generates roughly as fast as a small model while answering like a much larger one — provided it fits in memory. VRAM math here.

Check your card against a specific size: which GPUs run Qwen3 8B · which GPUs run Qwen3-Coder 30B.

Step 2 — Get it running

The fast path is Ollama. For a desktop GUI, LM Studio or open-source Jan let you search and download any Qwen size and chat immediately.

ollama
# Fastest path — Ollama
ollama pull qwen3:8b
ollama run qwen3:8b

# Coding variant for agents:
ollama pull qwen3-coder:30b

For full control of the GGUF and flags, use llama.cpp directly. Qwen handles long context well, so bump --ctx-size if your VRAM allows.

llama.cpp
# llama.cpp — pick the exact GGUF + quant
llama-server \
  --hf-repo Qwen/Qwen3-8B-GGUF \
  --hf-file Qwen3-8B-Q4_K_M.gguf \
  --n-gpu-layers 999 \
  --ctx-size 32768 \
  --host 0.0.0.0 --port 8080

Step 3 — Wire it into your coding agent

Qwen3-Coder really shines inside an agent. Any of them — Cline, Aider, Continue, Qwen Code — just needs an OpenAI-compatible base URL and a model name. We have step-by-step configs: Qwen Code, Cline, Aider, and Continue.dev.

Step 4 — Make it reachable from anywhere you code

The friction with local Qwen is the same as any local model: the runtime binds to localhostwith no auth, so it's only usable on the machine running it. If your GPU lives in the homelab but you code on a laptop — or you want the same model on your work machine and your teammate's — you need reach.

Wide Area Intelligence is a gateway that adds exactly that, without replacing your runtime:

01

An endpoint your editor can reach from anywhere

An outbound tunnel exposes the GPU box with no port forwarding, behind one URL (https://wideareaai.com/api/v1) and a revocable wai_sk_… key.
02

One config, many machines

Point Cline / Aider / Continue / Qwen Code at that endpoint once and it works from your laptop, your desktop, and a teammate's box — load-balanced across whichever GPU nodes are online.
03

Failover so the agent never hard-stops

If the node reboots mid-task, requests fail over to a cloud model on prepaid credits instead of erroring out.

Putting it together

Run Qwen3 8B for general use or Qwen3-Coder 30B for agents, pull it with Ollama or a GGUF in llama.cpp, confirm the fit with the GPU-for-model tool, and if you want that GPU reachable from every machine you work on, put it on the network with Wide Area Intelligence → Free for up to two nodes.

Related: How to run Llama 3 locally and How to run DeepSeek locally.

Frequently asked questions

What GPU do I need to run Qwen locally?
Qwen3 8B at Q4_K_M is about 5 GB of weights, so an 8 GB+ GPU runs it well. Qwen3 14B wants ~10 GB, Qwen3 32B around 20 GB (a 24 GB card). The Qwen3-30B-A3B mixture-of-experts model is unusually efficient — it has 30B total parameters but activates only ~3B per token, so it runs fast on modest hardware. Use the GPU-for-model tool to check a specific size.
Which Qwen model is best for coding?
Qwen3-Coder-30B and Qwen3-Coder-Next are purpose-built for code and are among the strongest open-weight coding models you can self-host. They pair well with coding agents like Cline, Aider, Continue, and Qwen Code. If VRAM is tight, the smaller Qwen2.5-Coder-7B is a capable fallback.
What is a Qwen MoE model and why does it matter locally?
Qwen3-30B-A3B is a mixture-of-experts (MoE) model: it has 30B total parameters but only activates about 3B per generated token. Because local single-user speed is bound by how much memory is read per token, an MoE model punches far above its total size — you get big-model quality at small-model speed, as long as it fits in memory.
How do I point my coding agent at a local Qwen model?
Run Qwen with any OpenAI-compatible server (Ollama, llama.cpp, vLLM) and point your agent at that base URL. To use it from a different machine or behind a real API key, put a gateway like Wide Area Intelligence in front — it gives a stable endpoint, revocable keys, an outbound tunnel, and cloud failover. See our Qwen Code and Cline guides for exact configs.

/// get started

That GPU is already paid for.
Put it on the network.

Create your gateway — free →